STATERA Library · Series 08 · Volume 15 · Chapter 11 · EN
Source: https://github.com/Statera-Library/08_Vol_15_Motion_Planning_and_Navigation/blob/main/Chapter_11_Legged_Robot_Navigation/Chapter_11_Legged_Robot_Navigation_EN.md
Volume 15. Motion Planning and Navigation
Chapter 11. Legged Robot Navigation
11.01. Legged Robot Navigation Unique Challenges and Opportunities
Legged robot navigation differs fundamentally from wheeled navigation because locomotion and navigation cannot be treated as fully independent problems. A legged robot must decide not only where to move but also how its body and individual feet should interact with the terrain. Every navigation decision therefore depends on terrain geometry, contact stability, body configuration, actuator limits, and the dynamic state of the robot.
The principal challenge is the discontinuous nature of ground contact. Unlike a wheeled platform that normally maintains continuous contact through its tires, a legged robot repeatedly creates and removes contacts while walking. The support region changes during each gait cycle, and disturbances can destabilize the body. Navigation must consequently coordinate global motion objectives with gait generation, foothold selection, balance control, and contact-force regulation.
Terrain traversability is substantially more complex for legged systems than for conventional mobile robots. A region that appears occupied in a two-dimensional costmap may actually be a useful stepping surface, while an apparently free region may contain holes, unstable material, excessive slopes, or geometric structures unsuitable for foot placement. Navigation therefore requires three-dimensional terrain representations that describe elevation, surface orientation, roughness, discontinuities, and expected contact quality.
Perception must provide information at several spatial scales. Long-range sensing identifies corridors, obstacles, slopes, stairs, and alternative routes, whereas short-range perception estimates precise geometry for upcoming footholds. LiDAR, depth cameras, stereo cameras, and inertial sensors are commonly combined to construct local elevation maps or other terrain models. The navigation system must continuously update these representations because body motion can cause significant sensor viewpoint changes.
Traversability estimation converts perceived geometry into information useful for planning. Instead of classifying space simply as free or occupied, the robot estimates whether a region can be crossed safely with its available locomotion capabilities. Factors such as slope angle, step height, surface roughness, foothold area, clearance, friction, and uncertainty may contribute to a traversability cost. This produces a mobility-aware representation of the environment.
Legged navigation naturally introduces multiple planning levels. A global planner can determine a relatively long-range route using terrain and mission information, while a local planner adapts motion to newly observed obstacles and terrain conditions. Below these layers, a locomotion planner generates body trajectories, gait transitions, and feasible footholds. Successful navigation requires these planners to exchange feasibility information rather than operating as completely isolated modules.
Foothold planning is one of the defining differences between legged and wheeled navigation. Candidate foot locations must provide sufficient support while respecting leg workspace, collision constraints, joint limits, and expected contact conditions. Selecting geometrically reachable footholds alone is insufficient because a sequence of individually valid contacts can still create an unstable or dynamically infeasible motion. Foothold evaluation must therefore consider the evolving robot state.
Dynamic stability becomes particularly important when the robot moves rapidly or traverses irregular terrain. Static measures such as the support polygon can be useful during slow motion, but dynamic locomotion requires consideration of momentum, contact forces, center-of-mass motion, and future contact transitions. Model Predictive Control (MPC) and related optimization techniques can predict these quantities over a finite horizon and continuously revise commands as new state estimates arrive.
State estimation is also more difficult because legged locomotion produces vibration, impacts, rapid orientation changes, and intermittent contacts. Inertial Measurement Unit (IMU) data can be fused with joint encoders, contact estimates, visual information, and LiDAR observations to estimate body pose and velocity. Incorrect assumptions about foot contact or slip can introduce significant errors, making robust contact estimation an important component of reliable navigation.
Foot slip and uncertain terrain create additional difficulties in outdoor environments. Mud, gravel, snow, vegetation, wet surfaces, and loose debris may have similar geometric appearances but very different mechanical properties. A robot may therefore need to combine geometric perception with proprioceptive observations such as unexpected foot velocity, contact force, actuator response, or body acceleration. Navigation can then modify speed, gait, footholds, or route selection.
The coupling between navigation and locomotion also creates an important opportunity: the robot can actively adapt its morphology and gait to the environment. It may lower its body beneath an obstacle, increase clearance over vegetation, shorten steps on uncertain terrain, widen its stance for stability, or select a different gait for higher speed. Configuration therefore becomes another planning variable rather than a fixed property of the mobile platform.
Legged robots can exploit terrain that would normally be considered an obstacle by wheeled systems. Stairs, curbs, rocks, narrow passages, discontinuous surfaces, and moderate gaps may become traversable when suitable footholds exist. This changes the meaning of obstacle avoidance. Instead of always planning around geometric obstacles, the system can evaluate whether stepping onto, over, between, or across environmental structures provides a safer or shorter route.
Navigation objectives must nevertheless balance mobility with risk. The shortest route may contain difficult terrain that increases energy consumption, slip probability, actuator loading, or fall risk. A more practical cost function can combine distance, traversal time, terrain difficulty, stability margin, energy demand, localization uncertainty, and recovery capability. Mission requirements determine how these costs should be weighted for inspection, logistics, exploration, or emergency-response applications.
Recovery behavior is especially important because navigation cannot assume that every planned motion will succeed. A foot may fail to establish secure contact, terrain may deform, or an unexpected obstacle may invalidate the planned foothold sequence. The robot should detect degraded conditions and transition to recovery actions such as stopping, widening its stance, replacing a foot, stepping backward, replanning locally, or requesting a new global route.
Learning-based methods provide opportunities to improve the relationship between perception, planning, and locomotion. Reinforcement Learning (RL) can produce terrain-adaptive locomotion policies, while learned traversability models can infer mobility characteristics from visual, geometric, and proprioceptive observations. These approaches can complement model-based planners, but deployment requires careful treatment of uncertainty, out-of-distribution terrain, physical constraints, and safety boundaries.
Semantic information can further improve navigation by allowing the robot to reason about environmental meaning rather than geometry alone. Stairs, doors, grass, pavement, rubble, ramps, and construction materials imply different locomotion strategies even when their local shapes are similar. Combining semantic perception with geometric terrain maps allows navigation costs and locomotion modes to reflect both physical structure and expected interaction properties.
For quadrupeds and other autonomous legged platforms, a robust architecture therefore connects perception, state estimation, terrain mapping, traversability analysis, global planning, local planning, foothold generation, locomotion control, and safety supervision in a continuous feedback loop. Information must flow upward as well as downward: locomotion failures and contact observations should update navigation-level estimates of what terrain is actually traversable.
The major opportunity of legged robot navigation is ultimately the ability to transform environmental complexity from a strict mobility limitation into a planning problem. By jointly reasoning about route, terrain, body configuration, contacts, and dynamics, a legged robot can operate where conventional platforms cannot. Achieving this capability requires navigation to evolve from two-dimensional collision avoidance into an integrated process of perception, prediction, contact planning, and adaptive physical interaction.
11.02. Footstep Planning for Rough Terrain Navigation [w/Code]
Footstep planning is a defining capability of legged robot navigation because the robot cannot move simply by following a continuous ground trajectory. Instead, it must determine a sequence of discrete contact locations for its feet while maintaining balance and satisfying the kinematic and dynamic constraints of the body. On rough terrain, this problem becomes tightly coupled with perception, terrain interpretation, body motion, gait selection, and local navigation.
A rough-terrain environment contains surfaces with different heights, slopes, gaps, rocks, steps, loose materials, and irregular boundaries. A useful terrain representation therefore needs more than an occupancy map. Elevation maps, surface normals, roughness estimates, traversability costs, and uncertainty information can be used to identify candidate footholds. The planner must continuously update this representation because newly observed terrain can invalidate previously selected footholds.
Footstep planning generally begins by generating candidate foothold locations from the local terrain model. A candidate should provide sufficient support area and acceptable surface orientation while remaining reachable by the corresponding leg. The planner also considers nearby obstacles, foot clearance, collision margins, joint limits, and the expected position of the robot body. Candidate generation should therefore be understood as a constrained geometric filtering process rather than simply selecting visually flat points.
Reachability is one of the most important constraints in foothold selection. Each leg has a limited workspace determined by its kinematic structure, joint ranges, current body configuration, and neighboring contacts. A foothold that appears safe from the terrain perspective may be unreachable without moving the body into an undesirable configuration. The planner must evaluate both individual leg reachability and the feasibility of the complete sequence of contacts that will move the robot toward its navigation goal.
Stability must be evaluated over the entire stepping sequence rather than at only one instant. During locomotion, the support polygon changes whenever a foot is lifted or placed, and the center of mass must remain compatible with the available contacts. For dynamic walking, the planner may additionally consider momentum, contact forces, body velocity, and predicted state evolution. This makes foothold planning closely related to whole-body motion planning and model predictive control.
The cost assigned to a foothold should represent more than geometric quality. A practical cost function can combine terrain roughness, surface slope, support area, distance from the nominal step, body stability, expected energy consumption, collision risk, and uncertainty. A slightly longer step onto a large and stable surface may therefore be preferable to a shorter step onto a narrow or uncertain surface. The appropriate weighting depends on whether the mission emphasizes speed, safety, endurance, precision, or terrain coverage.
Step sequencing introduces another level of complexity. Selecting four individually good footholds does not guarantee that the robot can safely transition between them. The sequence must respect the gait pattern, contact timing, leg workspace, body trajectory, and stability requirements. A planner may therefore search through alternative foothold sequences rather than optimizing each foot independently. This transforms foothold planning into a discrete-continuous planning problem involving both contact decisions and continuous body motion.
Body motion and foothold selection should be planned together whenever terrain becomes sufficiently difficult. The robot may need to shift its center of mass, rotate its body, change its height, or modify its stance width to reach a desired foothold. Conversely, the available footholds constrain the body trajectory that can be safely executed. Hierarchical planners commonly separate global route selection, local foothold planning, and low-level locomotion control while maintaining feedback between these layers.
Terrain uncertainty is particularly important in real environments. A depth camera or LiDAR may provide an apparently suitable surface while failing to reveal loose gravel, vegetation, mud, or an unsupported edge. The robot can reduce this uncertainty by incorporating proprioceptive feedback after each contact. Foot force, joint torque, estimated foot velocity, body acceleration, and contact confidence can indicate whether the actual terrain response matches the planned model.
Footstep planning must also account for dynamic obstacles and changing terrain. A foothold that was valid when initially selected may become unavailable because of another robot, a person, moving equipment, or newly revealed terrain. Local replanning should therefore operate continuously rather than treating the footstep sequence as a fixed trajectory. When conditions change significantly, the system can modify upcoming footholds while preserving already established contacts and maintaining a feasible transition.
Optimization-based methods provide a systematic framework for this problem. A planner can formulate foothold selection as an objective subject to terrain, kinematic, collision, stability, and timing constraints. Mixed-integer optimization, nonlinear programming, graph search, sampling-based planning, and model predictive approaches can all be applied depending on computational requirements. In practice, hierarchical combinations are often useful because global planning requires efficiency while local foothold decisions require detailed geometric reasoning.
Learning-based approaches can complement explicit foothold planners by estimating which terrain regions are likely to provide successful contacts. A learned model can use terrain appearance, geometry, previous contact outcomes, and proprioceptive information to predict foothold quality or slip probability. Reinforcement learning can also learn locomotion behaviors that implicitly adapt foot placement. However, learned predictions should remain bounded by explicit kinematic, collision, stability, and safety constraints when operating on physical robots.
The final foothold plan must be converted into executable locomotion commands. The locomotion controller determines swing trajectories, foot velocities, contact timing, body stabilization, and contact forces while tracking the planned sequence. During execution, deviations between the expected and measured robot state are fed back to the planner. This creates a closed-loop process in which planning proposes contacts, locomotion executes them, perception observes the result, and the next footholds are continuously revised.
For rough-terrain navigation, the most effective architecture is therefore not a simple foot-placement module but an integrated terrain-aware planning loop. Perception builds a three-dimensional terrain model, traversability estimation identifies suitable regions, foothold generation produces candidates, constrained planning selects feasible contacts, and locomotion control executes them while monitoring stability and contact quality. Navigation can then adapt its route when the terrain proves more difficult than initially expected.
The central objective of footstep planning is to transform irregular terrain from an uncontrolled obstacle field into a sequence of physically feasible decisions. A successful planner does not merely ask where a foot can be placed; it asks whether that placement contributes to a stable, reachable, collision-free, energy-efficient, and recoverable sequence of movements. This integration of terrain understanding, discrete contact planning, continuous body motion, and closed-loop feedback is what enables legged robots to navigate reliably beyond the limitations of conventional wheeled platforms.
11.03. Elevation Map Based Terrain Analysis [w/Code]
An elevation map is a fundamental representation for terrain-aware navigation because it describes the three-dimensional structure of the ground rather than treating the environment as a flat occupancy space. For a legged robot, elevation information is especially important because changes in height directly affect foothold selection, body posture, leg reachability, stability, and locomotion strategy. Elevation maps therefore provide an intermediate representation between raw perception data and motion planning.
An elevation map typically divides the surrounding terrain into a two-dimensional grid in which each cell stores an estimated ground height. The map can be constructed from LiDAR, depth cameras, stereo vision, or fused sensor observations. In addition to mean elevation, practical systems may maintain variance, observation count, surface normal, roughness, and temporal confidence. These additional attributes allow the planner to distinguish between well-observed stable terrain and geometrically uncertain regions.
The quality of an elevation map depends strongly on coordinate transformation and state estimation. Sensor measurements must be transformed from the robot or sensor frame into a consistent terrain frame using an accurate estimate of robot pose. Errors in roll, pitch, or vertical position can introduce artificial slopes and height discontinuities across the map. For legged robots, IMU measurements, joint states, visual odometry, LiDAR registration, and contact information can therefore be combined to maintain a stable terrain representation.
Unlike a static map, an elevation map for autonomous navigation should normally be local and continuously updated. As the robot moves, new terrain enters the sensor field of view while previously observed regions become less relevant. A rolling local map can maintain a fixed spatial area around the robot and update its cells using incoming observations. This approach limits computational cost while preserving the high-resolution terrain information required for local footstep planning and obstacle avoidance.
Terrain analysis begins by extracting geometric properties from the elevation data. Local gradients can estimate surface slope, while neighboring height differences reveal steps, edges, holes, and discontinuities. Surface normals provide information about the orientation of potential footholds, and local height variance can indicate roughness or measurement uncertainty. These geometric features transform a basic elevation map into a terrain description that can support navigation and locomotion decisions.
Slope is one of the most direct terrain characteristics derived from an elevation map. A moderate slope may be traversable, whereas a steep surface can exceed the robot\'s locomotion capability or create insufficient stability margins. However, slope should not be evaluated independently. A steep but broad and uniform surface may be safer than a nearly flat region covered with small unstable rocks. Terrain analysis should therefore combine slope with roughness, support area, contact geometry, and robot-specific mobility constraints.
Roughness estimation evaluates how irregular the terrain is within a local neighborhood. Height variance, residual error from a fitted plane, local curvature, or other geometric statistics can be used to characterize surface irregularity. Rough terrain does not automatically mean non-traversable terrain for a legged robot, because individual feet can often adapt to local height variations. The purpose of roughness estimation is instead to quantify the additional locomotion difficulty and provide useful information for foothold selection.
Elevation maps can also identify terrain discontinuities that are difficult to represent in conventional two-dimensional costmaps. Sudden height changes may indicate stairs, curbs, ledges, trenches, or large rocks. The planner can determine whether such structures should be avoided, crossed, stepped onto, or incorporated into a different locomotion strategy. This is particularly valuable for legged robots because vertical terrain structure can become a usable mobility feature rather than simply an obstacle.
Uncertainty is an essential component of elevation-map-based terrain analysis. Sparse LiDAR returns, reflective surfaces, vegetation, occlusion, sensor noise, and motion-induced distortion can produce unreliable elevation estimates. A planner should therefore distinguish between terrain that is known to be unsuitable and terrain whose suitability is simply unknown. Confidence-aware planning can avoid excessively optimistic decisions while preventing unnecessary avoidance of areas that may become traversable after additional observation.
Temporal information can further improve terrain interpretation. Multiple observations of the same region can be compared to determine whether a surface is stable or changing. A persistent elevation estimate provides greater confidence than a single measurement, while inconsistent observations may indicate moving objects, vegetation, deformable materials, or sensor artifacts. Temporal filtering must nevertheless be designed carefully so that rapidly changing obstacles are not incorrectly treated as stable terrain.
For legged navigation, elevation analysis should ultimately produce a traversability representation. Each terrain cell can be assigned a cost based on characteristics such as slope, roughness, elevation discontinuity, support area, clearance, uncertainty, and expected contact quality. Robot-specific constraints can then transform generic terrain geometry into mobility-aware information. The same elevation map may consequently produce different traversability costs for a quadruped, humanoid, wheeled robot, or tracked vehicle.
The elevation map also provides a foundation for foothold planning. Candidate foot locations can be sampled from cells with appropriate height, surface orientation, support area, and confidence. The planner can then evaluate whether each candidate is reachable by the corresponding leg and whether the resulting sequence maintains stability. This creates a direct connection between terrain perception and discrete contact planning, allowing the robot to use detailed local terrain geometry instead of relying only on a predefined global path.
Body motion can also be influenced by elevation-map analysis. When the terrain changes rapidly, the robot may adjust body height, pitch, roll, stance width, step length, or gait selection. A terrain map can therefore provide information not only about where the robot should place its feet but also about how the body should move while crossing the terrain. This enables coordinated planning in which route selection, foothold placement, and body configuration respond to the same environmental representation.
Elevation-map-based analysis becomes particularly powerful when integrated with semantic and learned terrain models. Geometry can determine the physical structure of a surface, while semantic perception can identify grass, pavement, stairs, gravel, mud, or other terrain categories. Learning-based models can combine appearance, geometry, and previous contact outcomes to estimate traversability or slip probability. The resulting representation can support more adaptive decisions while retaining explicit geometric constraints for physical safety.
A robust implementation must also consider computational efficiency and map resolution. High-resolution maps provide detailed foothold information but require greater memory, sensor-processing capacity, and planning computation. Coarser maps are efficient for longer-range terrain reasoning but may hide small obstacles or suitable stepping surfaces. A multi-resolution representation can therefore use coarse terrain information for broader navigation decisions and high-resolution elevation data for local foothold planning and locomotion control.
The complete process can be viewed as a continuous perception-to-action loop. Sensor observations are transformed into a consistent coordinate frame, fused into an elevation map, analyzed for slope, roughness, discontinuities, and uncertainty, and converted into traversability information. Navigation and footstep planners use this information to select routes and contacts, while locomotion execution generates new observations and proprioceptive feedback. These observations then update the map and trigger replanning when the actual terrain differs from the predicted model.
The central value of elevation-map-based terrain analysis is that it converts raw three-dimensional observations into a structured representation that can directly support physical decision-making. For a legged robot, the map is not merely a visualization of the ground surface; it is a computational model of where the robot can step, how difficult the terrain may be, and how its body should adapt. When combined with uncertainty estimation, traversability analysis, foothold planning, and closed-loop locomotion control, the elevation map becomes a core component of reliable rough-terrain navigation.
11.04. Stair and Step Detection and Traversal Planning [w/Code]
Stair and step detection is a critical capability for legged robot navigation because stairs and isolated steps create abrupt changes in terrain elevation that cannot be handled reliably by a conventional two-dimensional obstacle map. Unlike wheeled robots, legged robots can potentially climb, descend, or step across these structures, but doing so requires coordinated reasoning about geometry, foothold placement, body posture, leg reachability, balance, and gait. The navigation system must therefore recognize steps as structured terrain rather than simply classifying them as obstacles.
Detection begins by identifying elevation discontinuities in sensor observations. LiDAR, depth cameras, stereo cameras, and elevation maps can reveal vertical surfaces, horizontal tread regions, and repeated height transitions. Local surface normals, height gradients, edge features, and geometric consistency can be used to distinguish stairs from isolated rocks or irregular terrain. Robust detection should also tolerate partial occlusion, vegetation, shadows, missing depth measurements, and sensor noise because real environments rarely provide complete geometric observations.
An elevation-map representation is particularly useful because stairs have a characteristic spatial relationship between horizontal tread surfaces and vertical risers. A sequence of approximately uniform elevation changes can indicate a staircase, while a single abrupt transition may represent a curb or isolated step. The planner can estimate step height, tread depth, width, orientation, and spacing from the observed geometry. These parameters provide the information needed to determine whether the structure is compatible with the robot's locomotion capability.
Stair classification should not rely only on geometric appearance. The same elevation pattern may represent different physical situations, such as a concrete staircase, a damaged industrial platform, a rocky ascent, or a deformable surface. Semantic perception can provide additional context, while proprioceptive observations can confirm whether detected surfaces provide reliable support. Combining geometry, semantics, and contact information allows the robot to distinguish a visually recognizable step from a physically usable foothold.
After detecting a staircase or step, the robot must evaluate traversability. Step height is constrained by leg length, joint range, available torque, body clearance, and the desired gait. Tread depth determines whether a foot can be placed securely before the next transition, while staircase width affects the available lateral margin. Surface slope, roughness, friction, and uncertainty further influence whether the structure should be climbed, descended, bypassed, or approached using a different locomotion strategy.
Stair traversal requires planning a sequence of footholds rather than selecting a single target position. Each foot placement must remain reachable from the current body configuration while providing sufficient support for the next transition. The planner must consider contact order, swing-leg clearance, body trajectory, and the changing support polygon. A sequence that is individually valid at each step may still fail if the accumulated body motion creates an unstable configuration, making sequence-level planning essential.
Ascending and descending stairs impose different constraints. During ascent, the robot must generate sufficient vertical motion and maintain adequate forward progression while preventing the rear legs from colliding with risers. During descent, forward momentum and body pitch must be controlled carefully because the robot is moving toward lower terrain. The planner may therefore use different foothold margins, body trajectories, gait patterns, and speed limits for ascent and descent rather than treating both directions as identical problems.
Body posture adaptation is another important component of stair traversal. The robot may change body height, pitch, stance width, or step length to maintain a favorable configuration as elevation changes. A higher step may require greater leg extension, while a narrow tread may require more precise foot placement and a narrower motion envelope. Coordinating body motion with foothold planning reduces excessive joint excursions and helps maintain a stable relationship between the center of mass and the available support contacts.
The planner must also consider transitions between flat ground and stairs. The first step often represents a critical transition because the robot must shift from a conventional walking pattern to a terrain-adaptive sequence. The final step presents a similar challenge when returning to flat terrain. A robust system should therefore plan several contacts before and after the staircase rather than optimizing only the visible stair surfaces. This provides smoother transitions and reduces sudden changes in gait or body configuration.
Stair detection and traversal should operate within a hierarchical navigation architecture. Global planning determines whether the staircase is useful for reaching the mission goal, while local terrain analysis evaluates its actual geometric and physical condition. A foothold planner then generates feasible contacts, and the locomotion controller executes the selected sequence. If the observed staircase differs from the expected model, local replanning should modify upcoming steps without unnecessarily disrupting contacts that have already been established.
Uncertainty is especially important when stairs are partially visible or damaged. Missing tread edges, worn surfaces, debris, wet materials, and irregular risers can make geometric estimates unreliable. The robot should maintain confidence values for detected step boundaries and avoid making aggressive assumptions from incomplete observations. Additional sensing from a closer viewpoint can reduce uncertainty before committing to a difficult traversal. This creates an active perception loop in which navigation influences where the robot observes next.
Failure detection and recovery must be integrated into stair traversal planning. A foot may fail to establish the expected contact, slip on a tread, encounter an unexpected height difference, or collide with a riser. Contact force, joint torque, foot velocity, IMU response, and body motion can reveal such deviations. The robot can then stop, replace a foothold, modify the next step, change gait, retreat to a stable surface, or select an alternative route depending on the severity of the event.
Optimization and learning can further improve stair traversal. An optimization-based planner can minimize a combination of travel time, energy, instability, collision risk, and foothold uncertainty while satisfying kinematic and dynamic constraints. Learning-based models can estimate stair traversability, recognize unusual step structures, or adapt locomotion policies from previous experiences. In physical deployment, learned behavior should remain bounded by explicit safety and feasibility constraints so that unfamiliar stair geometry does not produce uncontrolled actions.
The complete stair-navigation process can therefore be understood as a closed-loop sequence of detection, geometric reconstruction, classification, traversability evaluation, foothold planning, body-motion adaptation, locomotion execution, and feedback-driven replanning. The essential objective is not merely to recognize that stairs exist, but to determine whether and how the robot can physically traverse them. By treating stairs and steps as structured three-dimensional terrain and coupling their analysis with contact-aware planning and locomotion control, a legged robot can turn abrupt elevation changes from navigation barriers into manageable mobility opportunities.
11.05. Jumping and Gap Crossing Motion Planning [w/Code]
Jumping and gap crossing introduce a fundamentally different class of locomotion problem because the robot temporarily loses ground contact and must control its body through a ballistic or highly dynamic motion phase. Unlike ordinary walking, where each step can provide continuous support and correction, a jump requires the robot to prepare its state before takeoff, control its motion during flight, and establish a stable landing afterward. Navigation and locomotion must therefore be planned as one continuous dynamic process.
Gap crossing begins with identifying whether a discontinuity can be crossed safely. Elevation maps, LiDAR, depth cameras, and stereo vision can estimate the width, depth, edges, and surrounding terrain of a gap. The planner must distinguish a traversable gap from a deep trench, unstable edge, or visually ambiguous region. Because perception may be incomplete, the estimated geometry should include uncertainty margins that prevent the robot from planning a jump based on overly optimistic measurements.
The key geometric variables include takeoff position, landing position, gap width, available approach distance, landing surface elevation, and the orientation of both terrain boundaries. A feasible jump requires sufficient space for the robot to generate the required takeoff motion and enough landing area to absorb the impact. The planner should therefore evaluate the entire region surrounding the gap rather than treating the gap width as the only relevant parameter.
Jump planning is strongly constrained by the robot\'s physical capabilities. Maximum takeoff velocity, leg workspace, actuator torque, body mass, available energy, flight time, and landing tolerance determine the maximum practical crossing distance. The planner should maintain explicit capability limits rather than relying solely on learned behavior. A geometric jump that appears feasible on a map may be dynamically impossible if the robot cannot generate sufficient vertical or horizontal velocity before leaving the ground.
The takeoff phase is particularly important because errors introduced before the robot becomes airborne cannot easily be corrected during flight. The robot may need to adjust its body height, stance width, center of mass, leg configuration, and approach velocity before initiating the jump. A suitable takeoff location should provide reliable footholds and sufficient traction so that the required impulse can be generated without slipping or rotating the body unexpectedly.
During takeoff, the locomotion controller must coordinate multiple legs to generate the required ground reaction forces. The desired impulse determines the initial velocity of the center of mass, while body orientation must remain within a range that permits a controlled landing. Excessive pitch or roll at takeoff can produce an unfavorable flight attitude and reduce the available correction margin. Model Predictive Control or whole-body optimization can be used to coordinate contact forces and body motion during this phase.
The flight phase differs from normal locomotion because the robot has no external ground contact through which it can directly regulate its motion. Body orientation can still be modified by moving the legs and redistributing internal angular momentum, but the available control authority is limited. The planner therefore needs to predict the future landing state from the takeoff condition and determine whether the robot will arrive with an acceptable position, velocity, and orientation.
Landing is often the most critical phase of a jump. The robot must establish contact with sufficient support while absorbing vertical and horizontal impact energy. Landing too close to an edge can cause partial support or rollover, while excessive horizontal velocity can produce slipping after contact. The landing controller should coordinate leg compliance, joint motion, body posture, and contact timing to distribute impact forces and rapidly recover a stable support configuration.
The landing surface should be evaluated using more than its apparent geometric area. Surface slope, roughness, height variation, friction, support strength, and nearby obstacles affect whether a landing is safe. A large but steep surface may be less suitable than a smaller flat region. For this reason, landing-site selection can be formulated as a foothold and terrain-quality problem in which multiple candidate landing configurations are compared according to stability and recovery margins.
Gap crossing also requires consideration of the approach trajectory. The robot may need to accelerate before reaching the takeoff point, which means the preceding terrain must provide enough distance for controlled speed generation. Obstacles or narrow corridors near the gap can restrict the approach angle and prevent the robot from achieving the desired takeoff state. Navigation planning must therefore reserve sufficient preparation space and cannot treat the jump as an isolated event.
The planner should distinguish between jumping over a gap and stepping across a small discontinuity. If a gap is sufficiently narrow, ordinary gait adaptation may provide a safer and more energy-efficient solution. The robot could lengthen a step, reposition its body, or select an alternative foothold sequence rather than entering a full airborne phase. Jumping should therefore be selected as a locomotion mode only when its expected benefit exceeds the additional dynamic and recovery risk.
Energy consumption is another important consideration. Dynamic jumping requires a substantial burst of actuator power and may increase mechanical stress compared with normal walking. Repeated jumps can reduce mission endurance and increase thermal or actuator loading. A navigation planner should therefore include energy and mechanical effort in its cost function and compare dynamic gap crossing with possible detours or slower terrain-adaptive walking.
Uncertainty becomes especially important because the consequences of an incorrect estimate are greater during a jump. A small error in gap width, landing elevation, friction, or takeoff velocity can produce a significant landing displacement. The planner should incorporate safety margins and confidence estimates into its feasibility calculation. When uncertainty is excessive, the robot may perform additional sensing, approach the edge more cautiously, select a wider landing region, or choose an alternative route.
Dynamic obstacle avoidance during jumping presents an additional challenge because the robot cannot stop or change direction normally once airborne. Moving objects near the landing zone can therefore invalidate an otherwise feasible plan. The navigation system should evaluate predicted obstacle motion and maintain a sufficiently large landing safety region. When necessary, the robot should delay the jump until the landing area becomes suitable rather than committing to an uncertain flight trajectory.
Jumping and gap crossing are well suited to hierarchical planning. A global planner determines whether crossing the gap is useful for the overall mission, while a local terrain planner identifies suitable takeoff and landing regions. A motion planner then generates the required body and contact trajectory, and the locomotion controller executes the dynamic maneuver. Feedback from perception and proprioception continuously updates the estimated state and can trigger replanning before takeoff or recovery actions after landing.
Simulation and predictive modeling are particularly valuable for dynamic gap crossing because many failures cannot be safely explored directly on a physical robot. Physics simulation can evaluate different takeoff velocities, body configurations, landing conditions, and terrain geometries. A digital twin or learned dynamics model can further estimate feasible jump envelopes. However, simulation results must account for model uncertainty, actuator limitations, contact variation, and the difference between simulated and real terrain.
Learning-based methods can complement model-based jump planning by learning takeoff and landing behaviors from repeated experience. Reinforcement learning can optimize coordinated leg motion, body orientation, and impact absorption, while supervised models can estimate successful landing regions from terrain observations. Nevertheless, safety-critical constraints such as maximum joint torque, allowable body orientation, collision clearance, and landing stability should remain explicitly enforced rather than delegated entirely to a learned policy.
Recovery planning must begin before the jump rather than being added only after failure. The robot should identify nearby stable terrain where it could recover if the intended landing is missed or partially successful. During flight, state estimation can determine whether the predicted landing remains feasible. If a safe landing is no longer possible, the controller may modify body orientation and leg configuration to maximize the probability of a controlled contact rather than continuing to follow the original trajectory.
The complete process can therefore be represented as a closed-loop dynamic navigation sequence: terrain perception identifies the gap, geometric analysis estimates its dimensions, feasibility analysis evaluates robot capability, navigation selects the crossing strategy, takeoff planning generates the required impulse, flight control manages body state, and landing control establishes stable contact. Proprioceptive feedback then verifies the outcome and updates the terrain and robot-state models for subsequent navigation.
The central objective of jumping and gap-crossing motion planning is not simply to maximize the distance a robot can jump. It is to determine when dynamic crossing is physically feasible, operationally useful, and sufficiently robust compared with alternative routes. By integrating terrain perception, capability modeling, takeoff and landing planning, dynamic control, uncertainty management, and recovery behavior, a legged robot can cross discontinuous terrain while maintaining the broader navigation objective and preserving safe, controllable mobility.
11.06. Full Body Path Planning with Height Maps [w/Code]
Full-body path planning for a legged robot extends conventional path planning from a simple body-center trajectory into a coordinated motion problem involving the entire robot body, legs, feet, and surrounding terrain. On uneven environments, a path that is geometrically collision-free for the robot\'s center may still be impossible to execute because the torso, knees, feet, or other body parts can collide with terrain. Height maps provide the geometric foundation for evaluating these interactions and generating physically feasible whole-body trajectories.
A height map represents terrain elevation over a two-dimensional spatial grid and can be augmented with surface normals, roughness, uncertainty, and traversability information. Full-body planning uses this representation to evaluate how the robot\'s complete geometry interacts with the terrain along a candidate path. Instead of asking only whether the robot can reach a destination, the planner determines whether the body can continuously move through the environment while maintaining sufficient clearance, feasible footholds, and stable configurations.
The first stage is to transform sensor observations into a consistent terrain representation. LiDAR, depth cameras, stereo vision, and IMU data can be fused to estimate terrain elevation while robot localization provides the transformation between sensor coordinates and the navigation frame. Because body motion can introduce significant perception distortion, the height map should be continuously updated using state estimation and temporal filtering. Uncertain or poorly observed regions should remain explicitly represented rather than being treated as reliably traversable.
Terrain clearance is a central consideration in full-body planning. A candidate body trajectory must maintain sufficient distance between the torso, legs, sensors, and other structures and nearby terrain. Low obstacles may be safely crossed by the feet but still interfere with the robot\'s torso or knee joints. Conversely, a narrow passage may appear blocked for the body center while remaining feasible if the robot changes its posture. Clearance therefore depends on the complete body configuration and cannot be represented by a single fixed footprint.
Body configuration becomes an important planning variable when terrain elevation changes significantly. The robot may lower its body under an overhead obstacle, raise its torso over rough terrain, rotate its body while traversing a slope, or widen its stance to improve stability. These changes affect leg reachability, joint limits, center-of-mass location, and collision margins. Full-body planning must therefore coordinate body pose with foothold placement rather than treating posture as a fixed parameter.
A candidate path is typically evaluated as a sequence of robot configurations distributed along a spatial trajectory. Each configuration includes the position and orientation of the body as well as the corresponding leg and foot states. The planner checks whether successive configurations can be connected through feasible motions while satisfying kinematic constraints. This converts ordinary path planning into a configuration-space problem in which terrain geometry and robot articulation are considered simultaneously.
Foothold planning remains closely coupled with the full-body path. Each foot must reach a suitable region of the height map while the body follows the planned trajectory. A body path that looks smooth may require impossible leg extensions or awkward joint configurations at particular terrain locations. Conversely, a carefully selected sequence of footholds can make a difficult body trajectory feasible. Whole-body planning therefore evaluates body motion and foot placement together whenever terrain complexity requires such coupling.
Dynamic stability adds another layer of constraints. A feasible geometric configuration is not necessarily dynamically stable during motion. The planner may need to consider the center of mass, support polygon, contact sequence, body velocity, angular momentum, and expected contact forces. For faster locomotion, model predictive control can evaluate predicted states over a finite horizon and continuously adjust the body trajectory and foothold sequence as the robot moves through the terrain.
Height maps can also support terrain-aware cost functions for whole-body path planning. A path may be penalized according to slope, roughness, body clearance, foothold quality, energy demand, collision risk, uncertainty, and expected locomotion difficulty. The shortest geometric route is therefore not necessarily the preferred path. A slightly longer route with smoother terrain and larger clearance may provide a more reliable and energy-efficient solution for the complete robot.
Multi-resolution planning is useful because different decisions require different spatial scales. A coarse height map can support long-range route selection and identify broad terrain structures, while a high-resolution local map can support precise body clearance and foothold decisions. The planner can therefore combine global terrain reasoning with local whole-body optimization without requiring the entire environment to be represented at maximum resolution.
Optimization-based methods provide a systematic way to solve the resulting constrained problem. The planner can minimize a combined objective involving path length, body motion, energy, stability, clearance, foothold quality, and collision risk while enforcing kinematic and dynamic constraints. Depending on computational requirements, graph search, nonlinear optimization, trajectory optimization, model predictive control, or sampling-based methods can be used. Hierarchical combinations are often practical because global navigation and local body-motion planning have different computational requirements.
Full-body planning must operate in a closed loop because the real terrain rarely matches the initial model perfectly. As the robot moves, new sensor observations may reveal unexpected steps, holes, obstacles, or surface changes. Proprioceptive feedback can also indicate foot slip, unexpected contact, excessive joint torque, or body-state deviations. The planner should use these observations to update the height map and modify upcoming body and foot trajectories rather than continuing to follow an outdated plan.
The central purpose of full-body path planning with height maps is to connect environmental geometry directly with physically executable robot motion. The height map provides terrain structure, while whole-body planning determines how the robot\'s body, legs, and feet should move through that structure. By integrating terrain perception, body clearance, foothold planning, posture adaptation, dynamic stability, optimization, and feedback-driven replanning, a legged robot can navigate complex three-dimensional environments with trajectories that are not merely geometrically valid but physically executable and robust.
11.07. Legged Robot Navigation in GPS Denied Environments [w/Code]
GPS-denied navigation is a critical capability for legged robots operating inside buildings, underground facilities, dense forests, urban canyons, tunnels, industrial plants, and other environments where reliable Global Navigation Satellite System (GNSS) signals are unavailable or degraded. Without an absolute global position reference, the robot must estimate its motion and construct a consistent local representation using onboard sensors. For legged robots, this problem is further complicated by body vibration, intermittent foot contact, terrain deformation, and dynamic locomotion.
A GPS-denied navigation system typically combines LiDAR, cameras, Inertial Measurement Units (IMU), joint encoders, and foot-contact information to estimate the robot state. Visual-Inertial Odometry (VIO) and LiDAR-Inertial Odometry (LIO) can provide continuous motion estimates, while leg kinematics and contact observations provide additional constraints. Sensor fusion is essential because every sensing modality has failure conditions. A robust system should therefore avoid depending on a single source of localization information.
State estimation is the foundation of navigation without an external positioning system. The estimator must continuously determine the robot\'s position, velocity, orientation, and sometimes sensor biases and terrain-related states. IMU measurements provide high-rate motion information but accumulate drift over time. Camera and LiDAR observations can reduce this drift through geometric registration, while contact constraints can provide useful corrections when a foot is reliably stationary relative to the terrain.
Legged robots provide a unique source of proprioceptive localization through their feet. When a foot establishes a stable contact, its estimated position can act as a temporary reference for body motion. Multiple contacts can constrain body translation and orientation, particularly during slow or quasi-static locomotion. However, contact assumptions become unreliable when the foot slips, the terrain deforms, or the robot encounters loose materials. Contact confidence should therefore be estimated explicitly and incorporated into the state estimator.
LiDAR-based localization is particularly valuable in visually degraded environments. A three-dimensional LiDAR can observe structural features such as walls, columns, machinery, terrain surfaces, and other persistent geometry. Scan matching or LiDAR odometry can estimate relative motion even when lighting conditions are poor. Nevertheless, geometrically repetitive environments can produce ambiguous matches, while dynamic objects can corrupt registration. Robust feature selection, motion compensation, and loop-closure mechanisms can reduce these problems.
Visual navigation provides complementary information in environments with rich textures and distinctive features. Cameras can detect structural edges, corners, semantic landmarks, and visual patterns that are difficult to represent using sparse geometric features alone. Visual-inertial estimation can operate at relatively high frequency and support precise local motion estimation. However, darkness, glare, dust, motion blur, repetitive surfaces, and low-texture environments can significantly reduce visual reliability, making sensor redundancy important.
The navigation system should maintain a local map even when a globally referenced map is unavailable. A local elevation map, voxel representation, point-cloud map, or semantic scene graph can provide the robot with spatial context for terrain interpretation and planning. The map should preserve geometric information needed for foothold selection, obstacle avoidance, and body-motion planning while also maintaining uncertainty. As the robot moves, local mapping and localization must operate together so that new observations remain correctly aligned.
Loop closure can provide an important mechanism for reducing accumulated drift. When the robot revisits a previously observed location, visual or LiDAR place recognition can identify the correspondence and introduce a global consistency constraint. Pose-graph optimization can then distribute accumulated error across the trajectory. For long-duration missions, loop closure is particularly valuable because even small localization errors can eventually distort the local map and cause the navigation system to make incorrect terrain or route decisions.
GPS-denied navigation also requires careful handling of localization failure. The robot should monitor indicators such as scan-matching residuals, visual tracking quality, inertial consistency, contact confidence, and estimated covariance. When localization quality deteriorates, the system can reduce speed, increase sensing frequency, select more observable terrain, or temporarily stop. Recovery may include relocalization against a stored map, returning to a previously known region, or switching to another available localization modality.
Terrain itself can become a localization resource. Large structural features, stairs, walls, corridors, poles, rocks, and other persistent geometric patterns can serve as landmarks. A legged robot can therefore combine navigation and terrain understanding rather than treating localization as an entirely separate function. In challenging environments, the planner may prefer routes containing distinctive and observable structures if they provide better localization confidence, even when several geometrically feasible routes are available.
Semantic perception can further strengthen GPS-denied localization. Objects and places such as doors, elevators, intersections, machinery, signs, storage racks, and building entrances can provide meaningful landmarks that remain useful across repeated missions. A semantic map can associate these landmarks with geometric locations and navigation context. This creates a richer localization framework in which the robot reasons not only about numerical coordinates but also about recognizable environmental structures.
Navigation planning should account for localization uncertainty rather than treating the estimated pose as perfectly known. A path through a feature-poor corridor may be geometrically short but produce significant positional uncertainty, while a slightly longer route containing strong geometric landmarks may maintain better localization quality. Cost functions can therefore incorporate distance, terrain difficulty, energy, obstacle risk, and localization confidence. This allows route planning to consider both where the robot should move and how reliably it can know where it is.
Multi-sensor fusion provides the strongest foundation for robust GPS-denied operation. An estimator can combine IMU, LiDAR, cameras, joint states, foot contacts, and occasionally external references such as UWB, radio localization, visual markers, or known infrastructure. The fusion architecture should allow individual sensors to degrade without immediately causing navigation failure. Measurement confidence, covariance estimation, outlier rejection, and fault detection are therefore important parts of the navigation stack.
For long-duration autonomous missions, map consistency and coordinate-frame management become increasingly important. The robot may maintain several coordinate frames, including a local body frame, odometry frame, local map frame, and mission-level reference frame. Transformations between these frames must remain consistent even when the robot experiences relocalization or loop closure. Clear separation between local motion estimation and higher-level global referencing helps prevent sudden navigation discontinuities.
The complete GPS-denied navigation process can therefore be understood as a continuous loop connecting perception, state estimation, mapping, localization, terrain analysis, planning, locomotion, and feedback. Sensor observations estimate the robot state and update the local map; localization confidence influences route and motion decisions; locomotion generates new observations and contact constraints; and detected inconsistencies trigger correction or recovery. The objective is not merely to replace GPS with another sensor, but to create a navigation architecture that remains reliable when absolute positioning is unavailable.
11.08. Hybrid Wheeled Legged Navigation Strategy
Hybrid wheeled-legged navigation combines the efficiency of rolling locomotion with the terrain adaptability of articulated legs. On smooth floors, roads, and gentle slopes, wheels can provide higher speed and lower energy consumption than repeated stepping. When the robot encounters stairs, large obstacles, gaps, rough terrain, or discontinuous surfaces, its legs can modify wheel position, body posture, or contact mode. Navigation therefore becomes a problem of selecting both a route and an appropriate locomotion mode.
The central advantage of a hybrid platform is that mobility does not need to be optimized around a single locomotion mechanism. Rolling can be treated as the preferred mode when terrain conditions permit because it minimizes mechanical motion and energy expenditure. Legged or articulated motion can then be activated only when necessary. This creates a mobility hierarchy in which the planner seeks the simplest feasible mode while retaining the ability to negotiate terrain that would block a conventional wheeled robot.
Terrain perception is essential for determining when locomotion modes should change. LiDAR, depth cameras, stereo vision, IMU data, and wheel or joint feedback can be fused into elevation maps, point clouds, or traversability representations. The system should identify surface slope, roughness, step height, obstacle dimensions, gaps, clearance, and uncertainty. These terrain characteristics are then evaluated against the capabilities of both rolling and leg-assisted locomotion.
A hybrid traversability map can represent more than a simple free-or-occupied classification. Each region may contain separate costs for rolling, stepping, climbing, body adjustment, or other supported locomotion modes. A flat corridor may have a very low rolling cost, while a staircase may have an effectively infinite rolling cost but a feasible legged cost. Navigation can therefore compare routes using mode-dependent terrain costs instead of assuming that every traversable region is equivalent.
Locomotion-mode selection becomes a discrete decision within continuous navigation. The robot may choose among rolling, wheel-assisted stepping, legged walking, obstacle climbing, posture adaptation, or stationary body reconfiguration. Each mode has different speed, energy, stability, clearance, and control requirements. The planner should select a mode according to terrain conditions and mission objectives while avoiding unnecessary transitions that increase execution time and mechanical complexity.
Mode transitions require explicit planning because switching from rolling to stepping is not instantaneous. The robot may need to reduce velocity, reposition its wheels, adjust leg extension, change body height, establish stable contacts, and confirm terrain conditions before entering a legged mode. Returning to rolling may similarly require aligning the wheels and restoring a suitable body configuration. Transition regions should therefore be considered part of the planned trajectory rather than treated as isolated controller events.
Body posture is particularly important for wheeled-legged systems because leg articulation can modify the effective geometry of the wheeled base. The robot can raise its body to increase ground clearance, lower it for stability, shift the center of mass on slopes, or adjust individual leg lengths to keep the body approximately level. These capabilities allow the planner to adapt the robot geometry to terrain rather than representing it with one fixed footprint and constant ground clearance.
Obstacle negotiation illustrates the benefit of hybrid locomotion. A small obstacle may be crossed by rolling directly over it, a larger obstacle may require lifting one or more wheels using leg articulation, and a very irregular region may require full stepping behavior. The navigation system should therefore classify obstacles according to the least complex safe traversal strategy. This reduces unnecessary dynamic motion while preserving the ability to handle challenging terrain.
Stairs create a particularly important mode-selection problem. Conventional rolling may be unsuitable, but wheel-assisted stepping can exploit both wheel contact and leg articulation. The robot may position wheels on successive treads while adjusting body pitch and leg extension to maintain stability. Stair traversal planning must consider tread depth, riser height, wheel radius, leg workspace, traction, and center-of-mass motion when deciding whether the structure can be crossed safely.
Slopes can often be traversed efficiently using wheels, but steep or irregular slopes may require active leg adaptation. The robot can modify individual leg lengths to maintain a favorable body orientation and redistribute normal forces across the wheels. If traction becomes insufficient, the planner may reduce speed, alter the approach direction, change posture, or switch to a more contact-intensive locomotion mode. Terrain slope and friction should therefore influence both path selection and mode selection.
Energy-aware navigation is especially valuable for hybrid robots because the energy characteristics of rolling and stepping are substantially different. Rolling generally requires less mechanical work on smooth terrain, whereas repeated leg motion can consume considerably more energy. A planner can include locomotion energy, actuator loading, transition cost, travel time, and terrain difficulty in its objective function. This allows the robot to choose a longer rolling route when it is operationally preferable to a shorter but highly articulated route.
Dynamic stability must be evaluated differently for each locomotion mode. During normal rolling, wheel contact and vehicle-like stability measures may dominate. During stepping, the support polygon, center-of-mass trajectory, contact forces, and leg reachability become more important. During transitions, both wheel dynamics and leg-contact constraints may apply simultaneously. A unified state representation can help the planner and controller manage these different stability regimes without treating them as unrelated systems.
Localization and mapping can also benefit from hybrid locomotion. Rolling provides smooth motion that can improve visual or LiDAR registration, while leg articulation allows the robot to reposition sensors or stabilize the body on irregular terrain. Conversely, rapid stepping or impact can introduce vibration and motion distortion. The navigation system can therefore consider sensing quality when selecting speed, posture, and locomotion mode, particularly in GPS-denied or visually difficult environments.
Hybrid planning is naturally suited to hierarchical architectures. A global planner can select a route based on distance, terrain classes, and estimated mode costs. A local planner evaluates detailed terrain geometry and chooses the immediate locomotion mode. A whole-body or contact planner determines wheel and leg configurations, while low-level controllers regulate wheel velocity, joint motion, contact forces, and body stability. Feedback between these layers allows feasibility information to influence higher-level navigation decisions.
Learning-based methods can improve mode selection by estimating terrain difficulty and predicting the success of different traversal strategies. Experience from previous missions can reveal that certain surfaces are more efficiently crossed by rolling, active suspension, or stepping. Reinforcement learning can optimize complex wheel-leg coordination, while supervised models can predict slip or traversal cost. Explicit safety constraints should nevertheless bound learned decisions when operating near stairs, gaps, unstable terrain, or other high-risk structures.
Failure detection is essential because a selected locomotion mode may become unsuitable during execution. Wheel slip, unexpected terrain deformation, joint saturation, excessive body tilt, or failed contact can indicate that the current strategy should be changed. The robot can reduce speed, adjust posture, switch from rolling to stepping, retreat, or replan the route. This capability turns locomotion-mode selection into a closed-loop process rather than a one-time decision made before traversal.
The complete hybrid navigation architecture can therefore be viewed as a continuous cycle connecting terrain perception, traversability analysis, route planning, mode selection, body-configuration planning, locomotion execution, and feedback. Each terrain segment is evaluated according to what the robot can accomplish using wheels, legs, or a coordinated combination of both. Execution results update terrain and mobility estimates, allowing subsequent decisions to reflect actual traction, stability, and energy performance.
The central objective of hybrid wheeled-legged navigation is not to use legs whenever terrain becomes difficult, but to select the most appropriate mobility mechanism for each situation. Wheels provide efficiency, speed, and smooth motion, while articulated legs provide adaptability, clearance, contact control, and access to discontinuous terrain. By integrating these capabilities within a unified planning framework, a hybrid robot can achieve broader mobility while controlling energy consumption, mechanical complexity, and operational risk.
11.09. Humanoid Navigation in Human Environments [w/Code]
Humanoid navigation in human environments requires more than moving a robot from one geometric position to another. A humanoid must operate in spaces designed around human body dimensions, behavior, and social conventions while maintaining balance through discrete foot contacts. Corridors, doors, stairs, elevators, furniture, crowds, and work areas therefore become both geometric constraints and interaction contexts that influence route selection, posture, walking speed, and motion execution.
The humanoid body provides important mobility advantages in human-centered infrastructure. Its narrow footprint can pass through pedestrian corridors, articulated legs can climb stairs and cross thresholds, and the upper body can change posture to negotiate constrained spaces. At the same time, a tall body and relatively small support region increase sensitivity to collisions and disturbances. Navigation must consequently consider full-body geometry rather than representing the robot as a simple circular or rectangular footprint.
Perception must capture both static environmental structure and dynamic human activity. LiDAR, RGB cameras, depth cameras, IMUs, and proprioceptive sensors can provide information about walls, doors, furniture, stairs, objects, people, and free space. Semantic perception can distinguish functional structures such as entrances, elevators, desks, shelves, and pedestrian areas. This information allows the navigation system to reason about how environmental elements should be used rather than treating every detected surface as generic geometry.
Human detection and tracking are particularly important because people move unpredictably and frequently change direction, speed, or intention. The robot should estimate pedestrian position and velocity while maintaining uncertainty about future motion. Short-term trajectory prediction can help identify potential conflicts before they occur. However, human motion should not be treated as perfectly deterministic, so the planner should preserve sufficient distance and maintain alternative actions when predictions become unreliable.
Social navigation introduces constraints that are not captured by collision avoidance alone. A humanoid should normally avoid passing unnecessarily close to people, cutting abruptly across walking paths, blocking doorways, or approaching someone from an uncomfortable direction. Personal space, pedestrian flow, group formations, and culturally dependent interaction conventions can influence navigation cost. These factors can be represented as soft constraints that modify route and velocity planning while physical safety remains a hard constraint.
Walking among people requires continuous velocity adaptation. A humanoid moving too quickly can create discomfort or reduce its ability to react, while excessive caution can obstruct surrounding pedestrians. The planner should adjust speed according to crowd density, visibility, available clearance, localization confidence, and predicted human motion. Smooth acceleration and deceleration are also important because abrupt changes affect both social predictability and the robot\'s dynamic balance.
Footstep planning connects pedestrian-scale navigation with humanoid locomotion. A global or local planner may generate a desired body path, but the robot must convert this path into a feasible sequence of left and right foot contacts. Each footstep must satisfy reachability, collision, stability, and terrain constraints. Narrow passages, turns, stairs, and cluttered areas may require modified step length, lateral stepping, body rotation, or temporary changes in walking pattern.
Whole-body collision avoidance becomes essential in confined environments. Even when the feet can follow a collision-free path, the shoulders, arms, torso, head, or carried objects may collide with surrounding structures or people. The planner should account for the changing volume occupied by the complete robot during walking. Arm configuration and torso orientation can become navigation variables, particularly when passing through narrow doors, moving beside furniture, or carrying large objects.
Doorways create a characteristic transition between open navigation and constrained full-body motion. The robot must detect the opening, estimate its width and orientation, align its body, select appropriate footsteps, and maintain clearance while passing through. If a door must be manipulated, navigation and manipulation become coupled because the robot must establish a suitable stance before operating the handle and then coordinate body motion with the changing door geometry.
Stairs are another major feature of human infrastructure. A humanoid can potentially use conventional stairs without specialized ramps, but stair traversal requires accurate estimation of tread depth, riser height, staircase orientation, and available handrails or support structures. Foot placement, body pitch, center-of-mass motion, and swing-foot clearance must be coordinated. Transitions at the first and final steps require particular attention because the locomotion pattern changes between level walking and vertical progression.
Elevator navigation combines locomotion with semantic reasoning and interaction. The robot must locate the elevator, approach an appropriate waiting position, avoid blocking exiting passengers, enter when sufficient space becomes available, and maintain balance during elevator motion. It may also need to interact with buttons or building systems. Successful operation therefore requires coordination among navigation, perception, task planning, communication, and potentially manipulation.
Crowded environments require local planning that can react rapidly without producing unstable or socially confusing behavior. The planner should preserve feasible stopping regions, avoid oscillating between alternative directions, and recognize when waiting is safer than forcing passage. In dense pedestrian traffic, cooperative behavior may emerge from yielding, following flow, or selecting a less congested route. Navigation quality should therefore consider progress, safety, predictability, and disruption to nearby people.
Localization in human environments can combine geometric, visual, and semantic information. Walls and structural features support LiDAR localization, while signs, doors, furniture, room identities, and other recognizable objects can provide semantic landmarks. Dynamic people should normally be separated from persistent mapping features. In GPS-denied buildings, multi-sensor localization and loop closure are especially important for maintaining map consistency across long corridors, multiple rooms, and different floors.
Navigation uncertainty should influence humanoid behavior directly. If localization confidence decreases, perception becomes partially occluded, or human motion becomes difficult to predict, the robot should reduce speed and increase safety margins. A narrow passage that is acceptable under accurate state estimation may become unsuitable under high uncertainty. Confidence-aware navigation therefore connects perception quality with locomotion aggressiveness and provides a systematic mechanism for conservative behavior when environmental understanding is incomplete.
Task context also changes navigation strategy. A humanoid delivering materials may prioritize efficient routes and stable carrying posture, while an inspection robot may intentionally approach equipment closely. A service robot interacting with people may prioritize visibility and socially appropriate approach direction. Navigation objectives should therefore include mission requirements in addition to distance, energy, terrain difficulty, collision probability, and social cost.
Human-aware navigation benefits from hierarchical planning. A mission-level planner selects destinations and semantic routes, a global planner determines paths through rooms and corridors, and a local planner manages people and nearby obstacles. Footstep and whole-body planners convert the selected motion into dynamically feasible locomotion, while low-level controllers maintain balance and contact forces. Feedback between these levels allows locomotion feasibility and social conditions to modify higher-level route decisions.
Learning-based methods can improve pedestrian prediction, semantic understanding, social navigation, and locomotion adaptation. Models trained from human motion data can estimate likely pedestrian trajectories, while reinforcement learning can optimize interaction-aware local navigation. Learned policies can also assist footstep generation and whole-body coordination. Physical collision limits, balance constraints, restricted areas, and minimum safety margins should nevertheless remain explicitly enforced during real-world deployment.
Failure detection and recovery are essential because human environments change continuously. A corridor may become blocked, an elevator may be crowded, a person may suddenly enter the robot\'s path, or a planned foothold may become unavailable. The robot should detect degraded conditions, stop safely when necessary, replan locally, select another route, or request assistance. Recovery behavior should be predictable to nearby people and should preserve stable locomotion rather than relying on abrupt evasive movements.
The complete humanoid navigation process can therefore be viewed as a closed-loop interaction among perception, semantic understanding, localization, human-motion prediction, route planning, social navigation, footstep generation, whole-body motion, and balance control. Environmental and human observations continuously update the robot\'s internal representation, while execution feedback modifies subsequent decisions. Navigation is consequently inseparable from locomotion and interaction in spaces shared with people.
The central objective of humanoid navigation in human environments is not simply to imitate human walking. It is to exploit a human-compatible body structure while providing reliable autonomous mobility within infrastructure and social spaces designed for people. By integrating semantic perception, human-aware planning, footstep and whole-body reasoning, dynamic balance, uncertainty management, and socially predictable behavior, humanoid robots can navigate complex shared environments safely, efficiently, and purposefully.
11.10. Quadruped Outdoor Terrain Navigation Case
A quadruped outdoor terrain navigation case represents an integrated Physical AI problem in which perception, localization, terrain understanding, route planning, footstep generation, locomotion, and safety control must operate continuously under changing environmental conditions. Unlike structured indoor navigation, outdoor deployment exposes the robot to slopes, rocks, vegetation, mud, loose soil, water, steps, gaps, and uncertain surfaces. Successful autonomy therefore depends on understanding both terrain geometry and its physical interaction with the robot.
Consider an autonomous quadruped assigned to inspect infrastructure across a large industrial or remote outdoor site. The mission may require traveling from a deployment station to several inspection points while crossing paved roads, gravel, grass, drainage structures, embankments, stairs, and unprepared terrain. The navigation system must select routes that satisfy mission objectives while considering travel time, terrain difficulty, energy consumption, localization confidence, communication coverage, and the probability of safe traversal.
The perception system combines three-dimensional LiDAR, RGB or depth cameras, an IMU, joint encoders, and proprioceptive contact information. Long-range sensing identifies large terrain structures and possible routes, while short-range sensing reconstructs detailed geometry for foothold planning. Sensor observations are transformed into a consistent local reference frame and fused into elevation maps, point clouds, or voxel representations. Continuous updates are necessary because vegetation, moving objects, shadows, and viewpoint changes can alter terrain interpretation.
Terrain analysis converts raw geometry into mobility-relevant information. Elevation gradients provide slope estimates, local height variation indicates roughness, and discontinuities reveal rocks, curbs, trenches, or steps. Surface normals and support-area estimates help identify candidate footholds. Uncertainty should remain explicitly represented because missing measurements or vegetation can make apparently continuous ground unreliable. The resulting terrain model becomes a robot-specific representation of traversability rather than a generic obstacle map.
Outdoor terrain requires more than geometric classification because visually similar surfaces can produce very different contact behavior. Dry soil, wet soil, gravel, grass, mud, and rock may have comparable shape but different friction, compliance, and load-bearing characteristics. Semantic perception can identify terrain categories, while proprioceptive feedback provides evidence about actual physical response. Combining exteroception and proprioception allows the robot to update its mobility estimate after interacting with the terrain.
Localization must remain robust when GNSS is degraded by trees, structures, terrain, or atmospheric conditions. GNSS can provide a global reference when available, while LiDAR-Inertial Odometry, Visual-Inertial Odometry, IMU integration, and leg-contact constraints maintain local motion estimates. Loop closure can reduce accumulated drift when the robot revisits known areas. Localization confidence should be communicated to the planner so that uncertain positioning leads to more conservative speed and route choices.
The global planner determines a mission-level route using terrain and operational costs rather than geometric distance alone. A short path over steep, loose terrain may be less desirable than a longer path over stable ground. Route cost can incorporate slope, roughness, predicted slip, energy consumption, localization quality, communication availability, and recovery difficulty. This enables the robot to select routes according to operational robustness instead of simply minimizing distance.
As the robot approaches difficult terrain, local planning becomes more detailed. The local planner evaluates nearby elevation and traversability information, modifies the body path, and determines whether normal walking remains appropriate. Step length, body height, gait, speed, and stance width can be adapted according to terrain conditions. If an obstacle cannot be crossed safely, the planner should identify an alternative local route and reconnect it to the mission-level path.
Foothold planning converts the selected local path into physically executable contacts. Candidate footholds are evaluated according to surface orientation, support area, roughness, reachability, collision clearance, and uncertainty. The planner must evaluate sequences rather than isolated contacts because each foothold changes the support geometry available for the next movement. Body trajectory and foot placement are therefore coordinated to maintain feasible leg configurations and adequate stability margins.
Slopes demonstrate the interaction between terrain analysis and locomotion adaptation. On moderate slopes, the robot can adjust body pitch and redistribute contact forces while maintaining its normal gait. On steeper or irregular slopes, it may shorten steps, lower the center of mass, widen the stance, or choose a diagonal route that reduces effective gradient. If estimated traction becomes insufficient, the robot should stop or replan rather than continuing until slip occurs.
Loose or deformable terrain requires closed-loop adaptation. A planned foothold may sink, slide, or provide less support than expected. Joint torque, contact force, foot velocity, body acceleration, and IMU measurements can reveal these conditions immediately after contact. The robot can then reduce speed, alter the gait, reposition the foot, redistribute load, or avoid similar terrain regions. The terrain model consequently evolves from purely visual geometry toward an experience-informed mobility representation.
Vegetation creates a particularly difficult perception problem because sensors may observe leaves and branches rather than the supporting ground. Some vegetation can be safely traversed, while dense bushes may conceal rocks, holes, or unstable surfaces. The planner should maintain higher uncertainty in such areas and use close-range sensing or cautious physical interaction before committing the full body. Semantic classification and previous traversal experience can improve decisions without eliminating explicit safety margins.
Water, mud, and drainage features introduce additional constraints. Reflective water surfaces may produce incomplete depth measurements, while mud can dramatically reduce traction and increase energy consumption. Drainage channels may appear as narrow gaps that can be stepped across or as deep structures requiring avoidance. The robot should evaluate width, depth uncertainty, landing quality, and alternative routes before attempting a dynamic or extended step across such features.
Outdoor obstacles may require locomotion-mode adaptation rather than simple avoidance. Rocks and fallen structures can sometimes be stepped onto or crossed, while low barriers may be negotiated by increasing foot clearance. Larger discontinuities may require specialized climbing or gap-crossing behaviors. The planner should compare these maneuvers against bypass routes using risk, energy, time, and recovery capability so that technically possible motions are not selected unnecessarily.
Weather and environmental change make long-duration navigation nonstationary. Rain can change friction, wind can disturb body stability, and changing illumination can degrade camera-based perception. Temperature can affect batteries and actuator performance. The navigation system should therefore incorporate environmental observations into capability estimates. A route considered safe earlier in the mission may require a different speed, gait, or path when conditions deteriorate.
Energy management is important when the quadruped must operate far from its deployment station. Terrain difficulty, gait, speed, payload, and repeated posture adjustments affect energy consumption. The mission planner should estimate the energy required not only to reach inspection points but also to return safely or reach a charging location. Navigation decisions can then trade distance against terrain efficiency, avoiding difficult shortcuts when they threaten mission endurance.
Communication loss must not immediately terminate autonomous navigation. The quadruped should retain sufficient onboard perception, mapping, planning, and locomotion capability to continue safely within predefined mission rules. When connectivity degrades, the robot may complete a local task, move toward a communication recovery point, or return along a known route. Mission-level autonomy should therefore distinguish between functions that require remote infrastructure and safety-critical functions that must remain available onboard.
Safety supervision operates independently of nominal navigation decisions. Excessive body tilt, joint load, localization uncertainty, slip, unstable contact, collision proximity, or degraded perception can trigger speed reduction, stopping, retreat, or replanning. Recovery locations and safe standing configurations should be considered during route planning. This prevents the robot from entering terrain where failure would leave no physically feasible way to stabilize or retreat.
The outdoor quadruped case can therefore be represented as a closed-loop hierarchy from mission planning to physical interaction. Mission objectives define destinations, global planning selects robust terrain corridors, local planning adapts the route, foothold planning determines contacts, and locomotion control executes body and leg motion. Perception and proprioception continuously evaluate the result, updating terrain, localization, traversability, and robot capability estimates before the next decision.
The essential lesson of quadruped outdoor navigation is that autonomy cannot be reduced to path following over a static map. Reliable operation emerges from continuous interaction between environmental understanding and physical mobility. By integrating terrain geometry, semantic interpretation, localization confidence, foothold planning, adaptive locomotion, energy management, safety supervision, and experience-based feedback, a quadruped can transform complex outdoor terrain into a sequence of measurable, executable, and recoverable navigation decisions.
Author: Richard Minsoo Go
LinkedIn: Richard Minsoo Go










Top comments (0)