STATERA Library · Series 08 · Volume 15 · Chapter 06 · EN
Source: https://github.com/Statera-Library/08_Vol_15_Motion_Planning_and_Navigation/blob/main/Chapter_06_Dynamic_Obstacle_Avoidance/Chapter_06_Dynamic_Obstacle_Avoidance_EN.md
Volume 15. Motion Planning and Navigation
Chapter 06. Dynamic Obstacle Avoidance
06.01. Dynamic Obstacle Avoidance Problem and Methods
Dynamic obstacle avoidance addresses the problem of navigating safely when obstacles change position over time. Unlike static path planning, where walls, shelves, and fixed structures can be represented by a stationary map, dynamic navigation must continuously reason about pedestrians, vehicles, forklifts, mobile robots, and other moving agents whose future states are uncertain. :chatgpt-content-reference{index="0"}
The fundamental difficulty is that collision risk depends not only on spatial separation but also on time. A location that is currently free may become occupied when the robot reaches it, while a currently blocked region may become available moments later. Dynamic obstacle avoidance therefore extends geometric collision checking into space-time reasoning, where position, velocity, heading, acceleration, and predicted motion influence every navigation decision.
A typical system begins with perception and tracking. Cameras, LiDAR, radar, depth sensors, or combinations of these sensors detect objects around the robot, while tracking algorithms estimate their positions and velocities over successive observations. The resulting dynamic world representation must distinguish moving objects from static structures and update rapidly enough that the planner does not operate on obsolete environmental information.
The robot then evaluates potential interactions between its own predicted motion and the estimated motion of surrounding objects. A simple approach assumes that each obstacle maintains its current velocity over a short horizon. More advanced approaches estimate multiple possible trajectories and their probabilities. Because human movement can change abruptly, prediction should normally be treated as uncertain rather than as a perfectly known future trajectory.
Collision assessment can be expressed through concepts such as time to collision, closest point of approach, predicted occupancy, or reachable regions. Instead of asking only whether two geometric footprints overlap, the system determines whether the robot and an obstacle could occupy conflicting regions at approximately the same time. Safety margins expand these regions to account for localization error, perception uncertainty, robot dimensions, braking distance, and prediction error.
Reactive avoidance methods generate commands directly from current or short-term observations. They are computationally efficient and can respond quickly to unexpected motion, making them useful as a fast local safety mechanism. However, purely reactive behavior may oscillate, become trapped, or select locally safe motions that create poor future situations. Consequently, practical navigation systems usually combine immediate reaction with short-horizon planning.
Velocity-space methods formulate avoidance by determining which robot velocities would produce a collision within a selected time horizon. The Velocity Obstacle concept identifies unsafe velocity regions generated by moving obstacles, allowing the controller to select admissible velocities outside those regions. Reciprocal formulations extend this principle to interactions in which multiple moving agents are expected to share responsibility for avoiding collisions.
Optimization-based methods instead evaluate candidate trajectories using objectives and constraints. A trajectory may be scored according to collision risk, progress toward the navigation goal, path deviation, control effort, smoothness, clearance, and dynamic feasibility. Model predictive approaches repeatedly solve a finite-horizon problem as new sensor measurements arrive, providing a natural framework for incorporating changing obstacle predictions into continuous navigation decisions.
Prediction-based avoidance becomes especially important when instantaneous reaction is insufficient. A pedestrian approaching an intersection may not yet block the robot\'s path, but trajectory forecasting can reveal an imminent conflict. The planner can then slow down, yield, stop, or pass before the interaction becomes critical. This predictive behavior generally produces smoother and safer navigation than waiting until the obstacle enters an immediate collision zone.
Dynamic obstacle avoidance also requires explicit consideration of robot dynamics. A commanded velocity cannot necessarily be reached instantaneously, and a heavy AMR may require substantial distance to decelerate. Acceleration limits, steering constraints, actuator delay, wheel-ground interaction, maximum curvature, and emergency braking capability therefore determine which avoidance maneuvers are physically executable rather than merely geometrically collision-free.
Human environments introduce additional requirements because mathematically collision-free motion may still be socially unacceptable. Robots should avoid passing excessively close to pedestrians, cutting across walking trajectories, approaching people aggressively from behind, or forcing humans into uncomfortable evasive actions. Human-aware navigation consequently incorporates interpersonal distance, directional preferences, yielding behavior, and other socially meaningful costs into planning.
Crowded environments create coupled interactions in which avoiding one person may increase the risk of collision with another. Local decisions must therefore consider several moving agents simultaneously. Methods based on reciprocal collision avoidance, social-force models, trajectory optimization, and learned interaction policies provide different mechanisms for handling these coupled behaviors, each making different assumptions about how other agents respond to the robot.
Safety-oriented methods can impose constraints independently of the nominal navigation objective. For example, a safety controller can restrict commands whenever the predicted state approaches an unsafe set, while a higher-level planner continues optimizing progress and efficiency. Control Barrier Function approaches formalize this concept by defining safety conditions that the controller attempts to preserve while modifying nominal control commands only when necessary.
Learning-based methods offer another approach by training policies from demonstrations, simulation, reinforcement learning, or recorded navigation data. Such policies can potentially capture complex interaction patterns that are difficult to encode analytically. Their main challenge is assurance: behavior learned from finite data may become unreliable in unfamiliar situations, so practical systems often combine learned prediction or planning with deterministic safety monitoring and fallback behavior.
No single avoidance method is universally optimal. Velocity-based techniques can provide efficient geometric reasoning, optimization methods can incorporate rich objectives and constraints, prediction-based approaches anticipate future conflicts, and learning methods can model complicated interactions. The chapter structure therefore treats these as complementary families, followed by specialized treatment of VO/RVO, ORCA, social-force models, prediction, CBF safety, and learned policies. :chatgpt-content-reference{index="1"}
A robust architecture consequently operates as a repeated perception-prediction-planning-control loop. Sensor observations update tracked objects, prediction estimates their future motion, collision analysis identifies dangerous interactions, the local planner generates feasible alternatives, and the controller executes the selected command. The complete cycle must run continuously because every executed motion changes the geometry and timing of subsequent interactions.
System quality should be evaluated beyond simple collision counts. Important measures include minimum separation distance, time to collision, intervention frequency, navigation success rate, travel time, path efficiency, stopping frequency, acceleration smoothness, computational latency, and social compliance. Testing should include crossing pedestrians, overtaking, head-on encounters, sudden appearances, narrow corridors, doorways, dense crowds, and interactions with other autonomous robots.
Dynamic obstacle avoidance is therefore not merely an obstacle-detection function but a real-time decision process connecting perception, prediction, motion planning, control, and safety. Within the broader motion-planning architecture, it bridges conventional local planning with human-aware navigation and multi-agent reasoning, providing the capability required for autonomous robots to operate safely and efficiently in environments that continuously change around them.
06.02. Velocity Obstacle VO and Reciprocal VO RVO [w/Code]
Velocity Obstacle (VO) is a geometric framework for dynamic collision avoidance that represents collision risk directly in velocity space rather than only in physical position space. Given the robot, a moving obstacle, their geometric footprints, and the obstacle velocity, VO identifies the set of robot velocities that would cause the two bodies to collide within a specified future interval. The avoidance problem then becomes selecting a feasible velocity outside this unsafe set.
The basic construction begins with relative motion. Instead of propagating the robot and obstacle independently through time, the obstacle velocity is subtracted from the robot velocity so that collision can be analyzed in a relative reference frame. The obstacle is conceptually stationary in this frame, while the robot moves according to the relative velocity. This transformation converts a dynamic two-body interaction into a simpler geometric collision test.
Robot and obstacle dimensions must also be incorporated into the calculation. A common approach uses the Minkowski sum to enlarge the obstacle by the robot footprint while representing the robot as a point. For circular approximations, the effective obstacle radius becomes the sum of the robot and obstacle radii. Collision checking can then be performed by determining whether the ray generated by the relative velocity intersects this enlarged obstacle region.
In velocity space, the resulting collision-producing velocities form a cone-like region whose apex and orientation depend on the obstacle velocity and relative position. Any robot velocity lying inside this Velocity Obstacle indicates that continuing with that velocity may eventually produce a collision. Velocities outside the region are considered collision-free under the assumed obstacle motion model and the selected prediction horizon.
The time horizon strongly affects VO behavior. An infinite horizon classifies every velocity that could eventually produce a collision as unsafe, which can make navigation unnecessarily conservative. A finite time horizon limits consideration to collisions expected within a practical interval. Short horizons encourage responsive and less conservative behavior, while longer horizons provide earlier avoidance but may restrict the available velocity space more aggressively.
When several moving obstacles are present, the robot constructs a Velocity Obstacle for each one. The union of these regions represents velocities that create predicted conflicts with at least one obstacle. The planner searches the remaining admissible velocity space for a command satisfying robot constraints while maintaining progress toward the goal. Desired velocity, clearance, smoothness, acceleration limits, and deviation from the reference path can be included when ranking candidates.
A major limitation of classical VO appears when both interacting agents independently perform avoidance. If two robots detect the same potential collision, each may change velocity in response to the other. At the next control cycle, both may observe that the conflict has shifted and reverse their decisions. Repeated symmetric reactions can generate oscillatory motion, especially during head-on encounters where both agents continuously choose corresponding avoidance directions.
Reciprocal Velocity Obstacle (RVO) was introduced to address this interaction by assuming that both agents participate in collision avoidance. Instead of requiring one robot to perform the entire avoidance maneuver, RVO distributes the required velocity change between the interacting agents. The resulting decision is based on the expectation that the other agent will make a complementary adjustment, reducing excessive reactions caused by treating its current velocity as permanently fixed.
Conceptually, RVO modifies the velocity-space construction using information from both agents\' current velocities. The reciprocal formulation shifts the effective avoidance region so that selecting a velocity outside it corresponds to taking an appropriate share of the collision-avoidance responsibility. When both agents apply compatible reciprocal reasoning, their combined velocity changes can separate their future trajectories without requiring either participant to execute the entire maneuver alone.
This reciprocal assumption is particularly useful in multi-robot environments. AMRs sharing warehouse intersections, service robots moving through corridors, and autonomous platforms operating in common traffic areas can encounter one another repeatedly. If each robot follows compatible avoidance rules, reciprocal methods can generate decentralized collision avoidance without requiring a central planner to calculate every short-term interaction among all robots.
However, RVO relies on assumptions about the behavior of other agents. A pedestrian, manually driven vehicle, or robot using a different navigation policy may not provide the expected reciprocal response. In such cases, assigning half of the avoidance responsibility to another agent can underestimate the maneuver required from the robot. Practical systems therefore combine reciprocal reasoning with conservative margins, obstacle classification, prediction, and safety overrides.
Velocity feasibility is another important consideration. Classical VO reasoning may identify a collision-free velocity that the physical robot cannot reach immediately. Differential-drive robots, Ackermann-steered vehicles, heavy AMRs, and dynamically constrained platforms have limits on acceleration, deceleration, turning rate, and curvature. The candidate velocity set should therefore be intersected with dynamically reachable velocities so that the selected avoidance action can actually be executed.
Uncertainty also changes the geometry of safe navigation. Position estimates contain localization and perception errors, measured velocities contain noise, and future obstacle motion rarely follows a perfectly constant-velocity model. Enlarging obstacle footprints, expanding unsafe velocity regions, shortening replanning intervals, or introducing probabilistic prediction can compensate for these uncertainties. The appropriate margin depends on sensing quality, platform dynamics, operating speed, and required safety level.
VO and RVO are naturally suited to real-time local navigation because their central computations are geometric and can be repeated whenever new tracking information arrives. A typical control cycle updates obstacle states, constructs velocity-space constraints, determines dynamically admissible robot velocities, removes collision-producing candidates, and selects a preferred safe command. This process is repeated continuously as both the robot and surrounding agents move.
The preferred velocity normally points toward a local navigation target or follows the global path, while collision constraints modify it only when necessary. This separation is useful because VO does not inherently determine where the robot should travel; it determines which velocities should be avoided. A local planner can therefore combine goal-directed motion with VO-based safety constraints while retaining responsibility for path following, progress, smoothness, and recovery behavior.
Dense scenes expose limitations of this approach because the union of many Velocity Obstacles may leave little or no feasible velocity space. Narrow corridors, doorways, opposing traffic, and tightly packed crowds can produce temporary states in which stopping or yielding is preferable to continuous avoidance. Deadlock detection, priority rules, trajectory prediction, global replanning, or explicit coordination may then be required above the basic VO or RVO layer.
The VO family provides an important bridge between geometric motion planning and multi-agent navigation. Classical VO expresses future collision risk as forbidden regions in velocity space, while RVO extends the idea by incorporating reciprocal behavior between moving agents. These principles establish the foundation for more refined reciprocal collision-avoidance methods such as Optimal Reciprocal Collision Avoidance (ORCA), which further formalizes feasible velocity selection for decentralized multi-agent motion.
06.03. ORCA Optimal Reciprocal Collision Avoidance [w/Code]
Optimal Reciprocal Collision Avoidance (ORCA) is a velocity-space method for decentralized collision avoidance among multiple moving agents. It extends the reciprocal reasoning introduced by RVO by converting pairwise collision-avoidance requirements into explicit linear constraints on admissible velocities. Each agent independently selects a velocity that remains collision-free over a specified time horizon while staying as close as possible to its preferred motion.
The central assumption of ORCA is reciprocal responsibility. When two cooperative agents are predicted to collide, neither agent is required to perform the complete avoidance maneuver. Instead, the minimum velocity correction needed to prevent the collision is determined, and each agent is normally assigned an equal share of that correction. This symmetric responsibility provides predictable interaction behavior without requiring centralized trajectory coordination.
ORCA begins with the same relative-motion geometry used by Velocity Obstacles. For two agents A and B, their relative position and relative velocity are evaluated together with their physical dimensions. The agents can commonly be approximated as discs in planar navigation, allowing their radii to be combined through a Minkowski-sum construction. Collision risk is then represented as a forbidden region in relative velocity space.
A finite time horizon defines which future collisions are relevant. The corresponding truncated Velocity Obstacle contains relative velocities that would cause the agents to collide before the horizon expires. If the current relative velocity lies inside this unsafe region, ORCA determines the smallest velocity displacement required to move it to the boundary of the collision-free region. This displacement becomes the basis for constructing the avoidance constraint.
The resulting ORCA constraint is represented as a half-plane in velocity space. Its boundary passes through a velocity derived from the current agent velocity plus its assigned fraction of the required correction. The boundary normal indicates the direction toward safe velocities. Any candidate velocity on the permitted side of this line satisfies the pairwise reciprocal collision-avoidance condition for the selected time horizon.
When an agent interacts with several neighbors, a separate ORCA half-plane is constructed for each relevant pair. The intersection of these half-planes forms a convex region containing velocities that satisfy all pairwise avoidance constraints simultaneously. The agent then selects a velocity within this feasible region that is closest to its preferred velocity, usually representing goal-directed motion generated by a higher-level navigation or path-following system.
This formulation turns multi-agent collision avoidance into a compact optimization problem. Because the constraints are linear in two-dimensional velocity space, feasible velocity selection can be solved efficiently using low-dimensional linear programming. This computational structure is one of ORCA\'s major advantages, enabling collision-avoidance decisions to be updated rapidly even when many agents are operating within the same environment.
The preferred velocity provides the connection between collision avoidance and navigation objectives. A global or local planner determines how the robot would ideally move toward its target if no dynamic conflict existed. ORCA does not replace this navigation objective. Instead, it modifies the preferred velocity only as much as necessary to satisfy the current reciprocal safety constraints, preserving progress and motion efficiency whenever sufficient free velocity space remains available.
ORCA is decentralized because every agent can compute its own constraints from locally observed neighboring states. It does not require a central controller to optimize the complete joint trajectory of the robot population. This property makes it attractive for multi-robot systems, warehouse AMRs, service robots, simulated crowds, and other applications where large numbers of autonomous agents must react quickly to nearby motion.
Compared with basic RVO, ORCA provides a more explicit mathematical definition of reciprocal avoidance responsibility. Rather than merely shifting an avoidance region according to reciprocal assumptions, it computes the minimum correction and defines a precise boundary separating admissible and inadmissible velocities. This produces a systematic velocity-selection procedure and makes the collision-avoidance problem suitable for efficient optimization.
The theoretical behavior depends strongly on the assumptions made about participating agents. Reciprocal guarantees are most meaningful when other agents also perform compatible avoidance actions and when their state estimates are sufficiently accurate. Humans, manually controlled vehicles, malfunctioning robots, or agents using incompatible policies may not contribute the expected share of avoidance, so practical robotic systems cannot always rely exclusively on reciprocal behavior.
Real robots also introduce dynamic constraints that are not fully captured by an instantaneous velocity-space model. A heavy AMR cannot arbitrarily jump from its current velocity to the velocity selected by ORCA. Acceleration, deceleration, steering rate, wheel configuration, actuator delay, braking distance, and curvature limits determine which commands are physically reachable. Velocity constraints should therefore be combined with platform-specific motion feasibility.
Perception and prediction uncertainty further affect practical deployment. ORCA calculations depend on estimates of neighboring positions, velocities, and dimensions, yet these quantities contain sensor noise and tracking errors. Obstacle inflation, conservative radii, uncertainty-dependent margins, shorter replanning intervals, and emergency safety layers can compensate for imperfect estimates and unexpected changes in surrounding motion.
The time horizon provides an important tuning mechanism. A longer horizon causes agents to react earlier to potential future conflicts and generally creates smoother anticipatory behavior, but it can also make the system more conservative in dense environments. A shorter horizon allows greater freedom of motion and later reactions, but may require stronger corrections when collisions become imminent. The appropriate horizon depends on speed, density, dynamics, and sensing latency.
Dense traffic can produce situations in which the intersection of ORCA constraints becomes extremely small or temporarily infeasible. Narrow corridors, doorways, intersections, and opposing robot flows are especially challenging because local reciprocal decisions cannot always resolve structural traffic conflicts. In these cases, stopping, yielding, priority assignment, reservation mechanisms, global replanning, or fleet-level traffic management may be required.
Static obstacles require different treatment because they do not share reciprocal responsibility. The robot must generally assume the full avoidance requirement for walls, shelves, pillars, parked equipment, and other non-cooperative objects. Practical navigation therefore combines reciprocal constraints for cooperative moving agents with non-reciprocal constraints for static or unresponsive obstacles, while maintaining a consistent feasible-velocity representation.
ORCA is particularly valuable as a local multi-agent avoidance layer rather than as a complete navigation architecture. Global planning can determine routes, fleet management can assign priorities, prediction can estimate future behavior, and safety controllers can enforce emergency constraints. ORCA operates within this larger system by efficiently transforming nearby agent interactions into velocity constraints that can be evaluated at every control cycle.
The method therefore provides an important progression from VO and RVO toward structured decentralized navigation. Velocity Obstacles identify collision-producing velocities, RVO introduces shared avoidance responsibility, and ORCA formalizes that responsibility through optimal velocity corrections and linear half-plane constraints. This combination of geometric reasoning, reciprocal interaction, and efficient optimization makes ORCA a foundational method for real-time multi-agent collision avoidance.
06.04. Social Force Model for Pedestrian Avoidance [w/Code]
The Social Force Model is a behavioral framework for representing pedestrian motion as the result of virtual forces that influence movement through shared space. Instead of treating people as passive moving obstacles, the model describes how pedestrians pursue destinations, maintain personal distance, avoid collisions, respond to boundaries, and interact with surrounding crowds. In robot navigation, these principles can be adapted to generate socially aware pedestrian avoidance behavior.
The central idea is that pedestrian motion can be approximated through a combination of attractive and repulsive influences. An attractive driving force moves a person toward a desired destination with a preferred velocity, while repulsive forces discourage excessive proximity to other pedestrians, robots, walls, and obstacles. The resulting acceleration is determined by combining these influences, producing continuous motion that resembles many common pedestrian interactions.
The driving force represents the tendency of an agent to approach its desired velocity. If the current velocity differs from the preferred velocity, the model generates an acceleration that reduces this difference over a characteristic relaxation time. The desired direction generally points toward a destination or waypoint, while the preferred speed reflects normal walking behavior. This formulation allows navigation intention and local interaction effects to be represented within the same dynamical model.
Repulsive social forces represent the tendency to preserve interpersonal separation. As two agents approach each other, the magnitude of the repulsive influence increases, encouraging them to alter their trajectories before physical contact occurs. The force can depend on relative distance, body dimensions, direction of motion, and model parameters describing personal space. Consequently, pedestrians begin avoidance before their physical boundaries intersect.
Interactions are generally anisotropic because humans respond more strongly to events occurring in front of them than to agents located behind them. A pedestrian approaching from the forward direction therefore produces a stronger avoidance response than an equally distant pedestrian behind the robot or person. Direction-dependent weighting can model this property and helps generate motion that better reflects visual attention and natural pedestrian behavior.
Static structures can be incorporated through repulsive boundary forces. Walls, shelves, pillars, curbs, and other environmental boundaries generate influences that discourage an agent from approaching them too closely. Combined with goal attraction, these forces guide movement through free space while maintaining clearance from surrounding structures. In narrow corridors, however, careful parameter tuning is required to avoid excessive repulsion that can prevent forward progress.
When the Social Force Model is applied to a mobile robot, the robot can be treated as another interacting agent whose motion is influenced by pedestrians and environmental boundaries. A desired-motion term pulls the robot toward its navigation target, while pedestrian-related forces modify the trajectory to preserve socially acceptable separation. This creates a continuous mechanism for balancing goal progress against local human avoidance.
The interaction can also be modeled asymmetrically. A robot may be configured to assume greater responsibility for maintaining separation than nearby pedestrians, producing more conservative behavior around humans. Different parameters can be assigned according to object type, allowing stronger avoidance around children, standing pedestrians, groups, or vulnerable users while applying different interaction rules to other autonomous robots or vehicles.
Pedestrian groups introduce additional complexity because people often move as social units rather than independent particles. Group members may attempt to remain close, walk side by side, preserve visibility among members, or maintain conversational formations. A robot that simply navigates through the geometric gap between group members may therefore behave socially inappropriately even if the trajectory is collision-free. Group-aware forces can discourage such intrusions.
The Social Force Model can also represent physical contact effects in extremely dense environments. When agents approach distances comparable to their physical dimensions, additional contact and friction terms can model compression and tangential interaction. These terms originated in crowd simulation applications but are generally less important for normal robot navigation, where the objective is to maintain sufficient separation so that physical contact never occurs.
One major advantage of the model is computational simplicity. Social interaction can be calculated from local relationships between the robot and nearby pedestrians without solving a large global trajectory optimization problem. Forces can be updated repeatedly as new observations arrive, making the approach suitable for real-time local navigation. The resulting control influence can also be integrated with conventional planners rather than operating as an independent navigation system.
However, social forces are mathematical abstractions rather than literal physical forces. Human decisions depend on intent, attention, culture, environment, social relationships, and contextual rules that cannot be fully represented through simple attraction and repulsion. Two pedestrians with identical positions and velocities may behave differently depending on whether they are crossing, waiting, conversing, following another person, or intentionally yielding to the robot.
Parameter selection therefore has a strong influence on navigation quality. Excessive repulsive strength may cause large detours, oscillation, or failure to pass through crowded areas, while weak repulsion can produce uncomfortable proximity. Relaxation time, interaction range, force magnitude, anisotropic weighting, preferred speed, and personal-space parameters should be calibrated using representative pedestrian scenarios rather than selected solely for geometric collision avoidance.
The method also differs fundamentally from Velocity Obstacle and ORCA approaches. VO-based methods reason explicitly about future collision-producing velocities, while the Social Force Model generates motion through continuous behavioral influences. Social forces can produce natural-looking interactions but do not inherently provide the same geometric collision-avoidance guarantees. Consequently, social-force behavior is often combined with explicit collision checking or an independent safety layer.
Prediction can further improve social-force navigation. Instead of calculating interactions only from current pedestrian states, predicted positions and trajectories can be used to estimate future social forces. A robot approaching a crossing pedestrian can then begin yielding before entering close proximity. Prediction is especially valuable when robot speed, braking distance, perception latency, or dense traffic makes purely reactive interaction insufficient.
Robot dynamics must also be respected when converting social forces into commands. The acceleration suggested by the behavioral model may exceed physical acceleration, steering, or braking capabilities. Practical implementations therefore transform the resulting desired motion into dynamically feasible velocity or trajectory commands. Saturation, smoothing, model predictive control, or local trajectory optimization can provide this interface between behavioral intention and physical execution.
Evaluation should consider social quality as well as conventional navigation performance. Collision rate and travel time remain important, but pedestrian clearance, intrusion into personal space, unnecessary stopping, path smoothness, yielding behavior, disturbance to pedestrian trajectories, and perceived comfort provide additional measures. Testing should include head-on encounters, crossing pedestrians, overtaking, groups, narrow corridors, doorways, queues, and dense bidirectional flows.
The Social Force Model therefore provides a useful bridge between dynamic obstacle avoidance and human-aware navigation. By representing goal attraction, interpersonal repulsion, environmental boundaries, directional sensitivity, and group behavior within a unified interaction model, it allows robots to reason about more than geometric collision. Its greatest value emerges when socially meaningful behavior is combined with prediction, dynamic feasibility, and explicit safety mechanisms within a complete navigation architecture.
06.05. Prediction Based Avoidance Object Trajectory Forecast [w/Code]
Prediction-based obstacle avoidance extends reactive navigation by estimating where moving objects are likely to be in the future rather than responding only to their current positions. The robot continuously observes pedestrians, vehicles, mobile robots, and other dynamic objects, predicts their future trajectories over a finite horizon, and evaluates whether those trajectories may intersect with its own planned motion.
The process begins with reliable object detection and tracking. Sensors such as cameras, LiDAR, radar, and depth cameras provide observations from which dynamic objects are identified and associated across consecutive frames. A tracker estimates state variables including position, velocity, acceleration, heading, and sometimes object class or behavioral state, producing the temporal information required for trajectory forecasting.
A simple trajectory predictor assumes constant velocity. Given the current position and estimated velocity of an object, its future position is propagated linearly through time. Constant-acceleration and constant-turn-rate models extend this representation when motion dynamics are more complex. These models are computationally efficient and often useful over short horizons, but their accuracy decreases when agents intentionally change direction or speed.
Model-based forecasting can incorporate kinematic and dynamic constraints appropriate to each object type. Vehicles may be predicted using steering and lane constraints, while mobile robots can be modeled according to differential-drive or Ackermann kinematics. Pedestrian prediction requires weaker physical assumptions because humans can rapidly change direction, stop, accelerate, or respond to surrounding agents, making behavioral context increasingly important.
Probabilistic trajectory prediction represents uncertainty explicitly rather than committing to a single future path. Instead of predicting one exact position at each future time, the predictor estimates distributions or confidence regions representing possible object locations. These regions often expand with prediction horizon because uncertainty accumulates over time. Collision checking can then consider both expected trajectories and their associated uncertainty.
Many dynamic interactions are inherently multimodal. A pedestrian approaching a corridor intersection may continue straight, turn left, turn right, or stop, while a vehicle may select among several feasible maneuvers. Multimodal forecasting represents multiple candidate futures together with their probabilities. This prevents the navigation system from assuming that the most likely trajectory is the only possible behavior and allows planning against several plausible outcomes.
Interaction-aware forecasting considers how agents influence one another. Pedestrians alter their paths to avoid other people, vehicles respond to surrounding traffic, and robots may yield or negotiate passage. Predicting each object independently can therefore produce inconsistent forecasts in dense scenes. Interaction models use neighboring states, relative geometry, social context, or learned relational representations to estimate mutually dependent future trajectories.
Scene context also constrains future motion. Walls, sidewalks, corridors, doors, lanes, crosswalks, intersections, and other semantic structures make some trajectories much more likely than others. Map-aware prediction incorporates geometric and semantic information so that forecasts remain consistent with navigable space. A pedestrian is more likely to follow a corridor than pass through a wall, while a vehicle normally remains within a drivable region.
Learning-based trajectory forecasting can capture complex patterns from recorded motion data. Recurrent networks, convolutional models, graph neural networks, transformers, and generative models can learn relationships between motion history, surrounding agents, and scene context. Modern predictors can generate multimodal trajectory distributions, but their performance depends strongly on training data coverage and may degrade when deployed in unfamiliar environments.
The robot must compare predicted object motion with its own candidate trajectories. Each candidate robot trajectory contains time-indexed states, and predicted obstacles also occupy time-dependent regions. A collision occurs when robot and obstacle occupancy overlap at corresponding future times. This space-time formulation is more informative than checking only spatial path intersections because two paths may geometrically cross while their agents pass through the intersection at different times.
Prediction-based avoidance therefore evaluates collision risk over a planning horizon. A trajectory can be rejected when predicted occupancy exceeds a safety threshold, or collision probability can be incorporated into a continuous cost function. Higher risk may cause the robot to slow down, increase clearance, yield, stop, or select an alternative path. The planner can balance this risk against progress, travel time, smoothness, and other navigation objectives.
Time-to-collision and time-to-conflict provide useful summaries of predicted interactions. An object that is physically close but moving away may require little intervention, while a more distant pedestrian rapidly approaching the robot\'s future path may require immediate attention. Forecasting therefore allows navigation decisions to reflect the temporal evolution of risk rather than relying exclusively on current Euclidean distance.
Prediction horizon selection creates an important tradeoff. Short horizons generally provide more accurate forecasts and reduce computational requirements but may detect conflicts too late for smooth avoidance. Longer horizons enable earlier anticipation but introduce greater uncertainty and may create unnecessary reactions to events that never occur. Practical systems often use a moderate horizon combined with frequent replanning as new observations become available.
Prediction uncertainty should propagate into the planning process. Treating forecast trajectories as exact can create unsafe confidence when an object deviates from its predicted motion. Safety margins can increase with prediction time, probabilistic occupancy can be integrated into collision costs, and low-confidence forecasts can trigger more conservative behavior. The planner should distinguish between highly predictable motion and ambiguous situations requiring additional caution.
Receding-horizon operation makes prediction-based avoidance robust to changing behavior. The robot does not calculate one forecast and follow the resulting trajectory indefinitely. Instead, sensing, tracking, forecasting, planning, and control are repeated continuously. As an object turns, stops, accelerates, or reveals its intent more clearly, new observations replace previous assumptions and the robot updates its planned motion accordingly.
Prediction alone cannot guarantee safety because unexpected behavior remains possible. A pedestrian may suddenly reverse direction, an object may emerge from an occluded region, or tracking may fail. Prediction-based planning should therefore operate together with short-range reactive avoidance, emergency braking, collision monitoring, or other independent safety mechanisms capable of responding when observed behavior differs substantially from the forecast.
Evaluation requires both prediction and navigation metrics. Forecasting quality can be measured using displacement error, final-position error, likelihood, calibration, and multimodal coverage, while navigation evaluation includes collision rate, minimum separation, success rate, travel time, path efficiency, braking frequency, and comfort. A predictor with low average trajectory error is not necessarily optimal if its rare failures create dangerous navigation decisions.
Prediction-based avoidance transforms dynamic obstacle handling from immediate reaction into anticipatory decision making. By combining tracking, trajectory forecasting, uncertainty representation, scene context, interaction modeling, space-time collision assessment, and continuous replanning, a robot can respond to where surrounding objects are likely to move rather than merely where they are now. This capability is fundamental for smooth and safe navigation in dynamic human and multi-robot environments.
06.06. Control Barrier Function CBF Safety Guarantee [w/Code]
Control Barrier Functions (CBFs) provide a mathematical framework for enforcing safety constraints in dynamical systems while allowing a robot to continue pursuing its nominal navigation objective. Instead of generating an entire motion plan, a CBF defines conditions that keep the system inside a designated safe set. A safety controller can therefore monitor a nominal command and modify it only when necessary to prevent the robot from entering an unsafe state.
The framework begins by describing robot dynamics in state-space form, where the state may contain position, orientation, velocity, or other variables relevant to motion. A control input determines how this state evolves over time. Safety is represented through a scalar barrier function whose value distinguishes safe states from unsafe states. Typically, the safe set is defined by states for which the barrier function remains nonnegative.
For obstacle avoidance, a barrier function can be constructed from the separation between the robot and an obstacle. A simple example compares squared distance with a required safety radius. When the robot is comfortably separated from the obstacle, the barrier value is positive. It approaches zero near the safety boundary and becomes negative after violation, providing a continuous mathematical representation of collision-related safety.
The essential CBF requirement constrains how rapidly the barrier function may decrease. It does not merely require the robot to be safe at the current instant; it restricts control actions so that the system cannot evolve across the boundary of the safe set. When the appropriate CBF condition is satisfied continuously, the safe set becomes forward invariant, meaning that a trajectory starting safely can remain within the safe region.
This concept distinguishes CBFs from ordinary collision penalties. A planner may assign a large cost to trajectories passing near obstacles, but optimization can still accept risk when other objectives compensate for that cost. A CBF instead introduces safety as a constraint that admissible controls are expected to satisfy. Navigation efficiency, path tracking, comfort, and energy consumption can then be optimized within the remaining safe control space.
A common implementation combines the CBF constraint with a quadratic program (QP). The nominal controller first produces a desired control command, such as linear and angular velocity or acceleration. The QP searches for the closest command that satisfies all active barrier constraints. When the nominal command is already safe, little or no modification is required; when danger increases, the safety filter alters the command sufficiently to preserve the safe set.
This architecture allows CBFs to operate as a minimally invasive safety filter. A global planner, local planner, reinforcement-learning policy, or human command can generate nominal behavior without explicitly solving every safety constraint internally. The CBF layer evaluates that command against current safety conditions and intervenes when required. This separation is particularly attractive when advanced or learned planners must be combined with a more interpretable safety mechanism.
Multiple obstacles can be handled by constructing separate barrier constraints for relevant objects. The resulting optimization problem requires the selected control input to satisfy these constraints simultaneously. Static walls, moving robots, pedestrians, vehicles, and restricted regions can therefore contribute independent safety conditions. Computational efficiency depends on the number of active constraints, so practical systems commonly prioritize nearby or relevant obstacles.
Dynamic obstacles require the barrier formulation to account for relative motion. Distance alone may be insufficient when an obstacle is moving rapidly toward the robot. Relative position, obstacle velocity, predicted motion, and robot dynamics can be incorporated so that the safety condition reflects how separation is changing over time. This enables the controller to begin braking or steering before the geometric safety boundary is physically reached.
For systems in which the control input does not directly affect the first derivative of the barrier function, higher-order formulations are required. This situation occurs when safety is defined using position but the robot controls acceleration rather than velocity. Higher-Order Control Barrier Functions (HOCBFs) recursively constrain derivatives of the safety condition, allowing acceleration-controlled and dynamically complex systems to enforce position-related safety requirements.
CBFs can also be combined with Control Lyapunov Functions (CLFs). A CLF represents progress or stability toward a desired objective, while a CBF represents safety. A joint CLF-CBF quadratic program can therefore seek control actions that advance toward a target while respecting safety constraints. If progress and safety conflict, the formulation can assign safety higher priority and relax the performance objective rather than violate the protected region.
The quality of a CBF safety guarantee depends on the accuracy of the underlying system model and state information. Localization errors, delayed perception, uncertain obstacle velocities, actuator lag, and unmodeled dynamics can invalidate assumptions used in the barrier constraint. Practical implementations therefore require margins that reflect sensing uncertainty, control latency, braking capability, and modeling error rather than treating estimated states as perfectly accurate.
Robust and probabilistic extensions can address uncertainty more explicitly. Robust CBF formulations account for bounded disturbances or model errors when constructing admissible controls, while uncertainty-aware approaches can incorporate confidence regions around obstacle states. These methods generally become more conservative as uncertainty increases, trading some navigation efficiency for stronger protection against estimation and prediction errors.
Feasibility is another important issue. Several simultaneous barrier constraints may demand mutually incompatible control actions, particularly in dense crowds, narrow passages, or situations where the robot is already moving too quickly to stop safely. Input limits can further restrict the solution. Safety architecture must therefore consider constraint prioritization, braking policies, emergency stops, and planning strategies that avoid entering states from which no feasible safe action remains.
CBFs are especially valuable when integrated with predictive planning. A trajectory predictor or local planner can anticipate future interactions, while the barrier layer protects against short-term deviations and unexpected behavior. Prediction provides smoother early avoidance, whereas the CBF provides a mathematically structured last line of control modification. The two mechanisms therefore address different temporal aspects of dynamic obstacle safety.
Real-time performance is critical because the safety filter must execute at the control frequency. The QP should remain sufficiently small and deterministic for the target computing platform, and sensor-to-control latency must be considered when defining safety margins. Efficient constraint selection, warm starting, bounded solver time, and fallback commands are important engineering considerations when deploying CBF-based control on physical robots.
Testing should include more than nominal collision-free scenarios. Relevant cases include sudden pedestrian crossings, rapidly approaching objects, sensor delay, noisy tracking, actuator saturation, emergency braking, narrow corridors, multiple simultaneous obstacles, and infeasible constraint combinations. Evaluation should measure minimum separation, barrier values, intervention frequency, control deviation, solver latency, navigation success, and the occurrence of safety violations.
A Control Barrier Function should therefore be viewed as a safety-enforcement mechanism within a larger navigation architecture rather than as a complete planner. Perception estimates the environment, prediction anticipates dynamic behavior, planning proposes efficient motion, and the nominal controller generates commands. The CBF safety filter constrains those commands so that safety conditions are maintained whenever the modeling assumptions and feasible control authority permit.
CBFs provide an important transition from heuristic obstacle avoidance toward formally structured safety control. By defining safe sets, enforcing forward-invariance conditions, and projecting nominal commands into an admissible control region, they allow performance-oriented navigation and safety enforcement to coexist. For dynamic robot navigation, this makes CBFs particularly useful as a supervisory layer connecting advanced planning or learned policies with physically executable safety constraints.
06.07. Learning Based Dynamic Avoidance Policies [w/Code]
Learning-based dynamic avoidance policies use data-driven models to select robot actions in environments containing pedestrians, vehicles, mobile robots, and other moving obstacles. Instead of defining every avoidance rule analytically, the policy learns relationships between observations, surrounding motion, navigation goals, and appropriate control actions. This allows complex interaction patterns to be represented within a unified decision-making model.
The learning problem can be formulated as a mapping from observations or estimated states to actions. Inputs may include robot pose, velocity, goal direction, LiDAR scans, camera features, occupancy representations, tracked-object states, or predicted trajectories. Outputs may directly specify linear and angular velocity, steering commands, acceleration, motion primitives, local waypoints, or parameters used by a conventional motion planner.
Imitation learning trains avoidance behavior from demonstrations provided by human operators, expert planners, or previously validated controllers. The model learns to reproduce actions associated with observed navigation situations. This approach can generate useful behavior without manually designing a reward function, but its performance depends strongly on demonstration quality and coverage. Rare dangerous situations may be poorly represented in ordinary navigation datasets.
Behavior cloning is the simplest imitation-learning approach, where supervised learning predicts expert actions from recorded observations. A major difficulty is distribution shift: small policy errors can move the robot into states that were rarely encountered in the demonstrations, causing further errors to accumulate. Interactive data collection and corrective demonstrations can reduce this problem by adding examples from states visited by the learned policy itself.
Reinforcement learning (RL) instead trains a policy through interaction with an environment. The robot receives rewards for desirable outcomes such as goal progress, collision avoidance, smooth motion, and efficient travel, while penalties discourage collisions, excessive proximity, oscillation, or unnecessary control effort. Through repeated experience, the policy learns action strategies that maximize expected cumulative reward rather than simply reproducing an expert controller.
The reward function strongly shapes learned behavior. A large collision penalty can encourage safety but may also create excessively conservative policies that stop whenever uncertainty appears. Strong progress rewards can improve travel efficiency but may produce aggressive interactions with pedestrians. Practical reward design therefore balances safety, progress, clearance, smoothness, comfort, energy consumption, and socially appropriate behavior while avoiding unintended optimization shortcuts.
Dynamic avoidance is naturally modeled as a partially observable decision problem because the robot cannot directly know another agent\'s intent. Current sensor measurements reveal only part of the environment, while hidden variables may include destination, attention, willingness to yield, or future maneuver. Recurrent neural networks, temporal encoders, belief representations, and trajectory histories can provide memory that helps the policy infer relevant motion patterns from sequences of observations.
Multi-agent reinforcement learning extends this formulation to environments containing several interacting decision makers. Robots can learn policies that account for the reactions of neighboring agents rather than treating them as independently moving obstacles. Training may be centralized while execution remains decentralized, allowing richer interaction information during learning while enabling each deployed robot to make decisions using locally available observations.
Graph neural networks (GNNs) provide a useful representation when the number of surrounding agents changes over time. The robot and nearby pedestrians or robots can be represented as nodes, while relative position, velocity, visibility, or interaction relationships form edges. Message passing allows the policy to aggregate information from relevant neighbors and learn relational patterns that support scalable multi-agent collision avoidance.
Attention and transformer-based models offer another mechanism for selecting which agents and temporal observations matter most. Instead of treating all surrounding objects equally, attention can emphasize pedestrians approaching the robot\'s path, vehicles with high closing velocity, or agents near an intersection. Temporal attention can also connect current observations with earlier motion states, supporting more context-sensitive decisions in complex dynamic scenes.
Learning can operate at different levels of the navigation architecture. An end-to-end policy may map raw sensor observations directly to control commands, while a modular architecture may use learning only for pedestrian prediction, interaction modeling, cost estimation, local trajectory ranking, or waypoint selection. Modular designs generally provide clearer interfaces and make it easier to retain conventional planning and safety components around the learned function.
Simulation is particularly important because collecting dangerous interaction data on physical robots is costly and risky. Large numbers of pedestrian crossings, head-on encounters, overtaking events, dense crowds, and multi-robot conflicts can be generated synthetically. Domain randomization can vary sensor noise, obstacle behavior, robot dynamics, appearance, latency, and environmental geometry so that the learned policy encounters a wider range of conditions during training.
The simulation-to-reality gap remains a major deployment challenge. A policy may exploit simplified simulator behavior that does not exist in the physical world, while real sensors introduce noise, occlusion, timing variation, and tracking failures. System identification, realistic sensor simulation, domain adaptation, randomized dynamics, and limited real-world fine-tuning can reduce this gap, but extensive physical validation remains necessary before safety-critical deployment.
Generalization is equally important. A policy trained in one building, crowd density, robot platform, or pedestrian behavior distribution may perform poorly in unfamiliar environments. Training across diverse maps, interaction patterns, speeds, sensor conditions, and agent populations improves robustness. Out-of-distribution detection can also identify observations that differ substantially from training experience and trigger conservative fallback behavior.
Learning-based policies should not be assumed to provide safety guarantees merely because training produced a low collision rate. Neural policies can fail unpredictably in rare states, and empirical success does not establish formal invariance or guaranteed braking capability. A practical architecture can therefore place deterministic collision monitoring, emergency stopping, Control Barrier Functions, or other verified safety filters around the learned policy.
A hybrid architecture allows learning and model-based methods to complement each other. A neural network may predict pedestrian intent or generate a preferred local motion, while VO, ORCA, optimization, or CBF mechanisms enforce geometric and dynamic constraints. Alternatively, learning can tune planner costs or select among safe candidate trajectories. Such combinations preserve the adaptability of learned behavior while retaining interpretable constraints and established navigation structure.
Real-time deployment imposes additional engineering constraints. Neural inference, perception, tracking, planning, and safety filtering must fit within the control-cycle latency budget. Model size, accelerator availability, memory bandwidth, worst-case inference time, sensor synchronization, and communication delay can affect navigation safety. Average inference speed alone is insufficient when occasional long latency causes the robot to react too late to a moving obstacle.
Evaluation should therefore extend beyond average task reward. Important measures include collision rate, minimum separation, navigation success, time to goal, path efficiency, intervention frequency, control smoothness, pedestrian disturbance, social compliance, inference latency, and robustness under sensor or behavior perturbations. Rare-event testing and adversarial scenarios are especially valuable because safety failures may be hidden by strong average performance.
Learning-based dynamic avoidance ultimately provides a flexible mechanism for representing interactions that are difficult to encode using fixed analytical rules. Its strongest role is not necessarily to replace classical navigation, but to contribute learned prediction, relational reasoning, adaptive decision making, or local policy generation within a layered architecture. Combined with explicit dynamic constraints and independent safety mechanisms, learned policies can support more adaptive navigation in complex changing environments.
06.08. Avoidance in Narrow Corridors and Doorways [w/Code]
Avoidance in narrow corridors and doorways is a constrained navigation problem in which limited free space reduces the robot's ability to maneuver around dynamic obstacles. Unlike open environments, lateral escape directions may be unavailable, and small deviations can place the robot close to walls, door frames, pedestrians, or other robots. Successful navigation therefore requires coordinated reasoning about geometry, timing, priority, motion feasibility, and interaction behavior.
The first requirement is accurate representation of traversable width. The planner must consider the robot footprint, localization uncertainty, wall clearance, sensor error, and additional safety margins rather than relying only on the geometric centerline of the corridor. A passage that appears open on a point-based map may be physically infeasible for a wide platform once its complete footprint and required clearance are considered.
Doorways create particularly strong geometric bottlenecks because the available width changes abruptly. The robot may approach through a relatively open region and then enter a passage only slightly wider than its footprint. The planner should therefore evaluate not only whether the doorway can be crossed, but also whether the robot can align itself before entry and maintain an appropriate orientation throughout the constrained transition.
Footprint orientation becomes important for noncircular robots. A rectangular AMR may fit through a doorway when aligned with the passage but violate clearance constraints while rotating near the entrance. Collision checking must therefore consider the swept footprint rather than evaluating only discrete robot poses. Entry trajectories should position and orient the platform before it reaches the most restrictive part of the passage.
Dynamic obstacles make these geometric constraints more difficult because two agents may be individually capable of traversing a corridor but unable to pass each other safely. A pedestrian or another robot approaching from the opposite direction can create a temporary conflict in which continuous forward motion is impossible. The navigation system must recognize that slowing, stopping, yielding, or retreating may be safer than attempting local side-to-side avoidance.
Passing feasibility can be evaluated from corridor width, agent dimensions, required separation, and uncertainty margins. If sufficient lateral space exists, both agents may pass while maintaining reduced speed and controlled clearance. If the available width is below the required combined envelope, the problem becomes one of temporal coordination rather than geometric avoidance. One participant must normally wait outside or move toward a wider region.
This distinction is important because conventional local avoidance algorithms may continue searching for a collision-free velocity even when no physically meaningful passing maneuver exists. Such behavior can produce oscillation, repeated left-right commands, or unnecessary proximity to walls. Detecting corridor topology and bottleneck conditions allows the planner to switch from continuous avoidance to explicit yielding or passage-management behavior.
Waiting locations should be selected deliberately. Stopping directly inside a doorway can block traffic in both directions, while stopping too close to a narrow entrance may prevent another agent from exiting. A robot can instead identify waiting zones, corridor widenings, alcoves, intersections, or staging positions where it can safely yield without obstructing the constrained region. These locations can be encoded in maps or detected from local free-space geometry.
Priority rules determine which agent should enter a bottleneck first when simultaneous passage is impossible. Priority may depend on arrival time, distance to the doorway, mission urgency, robot type, direction of traffic, loaded state, or fleet-level scheduling. Simple deterministic rules are often preferable to ambiguous local negotiation because two agents independently deciding to yield can create deadlock just as easily as two agents simultaneously deciding to proceed.
Reservation mechanisms provide a stronger form of coordination for multi-robot systems. A doorway, narrow corridor segment, elevator entrance, or other bottleneck can be represented as an exclusive or capacity-limited resource. Before entering, a robot requests permission or reserves a time interval. Other robots remain outside until the resource becomes available, reducing head-on conflicts and preventing several robots from entering a region that cannot support safe passing.
Human interactions require different assumptions because pedestrians do not participate in robot reservation protocols. The robot should infer pedestrian motion and preserve enough space for natural human behavior. Near a doorway, people may stop unexpectedly, emerge from occluded areas, change direction, or hold the door for another person. Conservative speed, trajectory prediction, and increased attention to blind regions are therefore important when approaching human-shared bottlenecks.
Occlusion is especially significant around door frames, corners, shelves, and corridor intersections. A pedestrian may become visible only shortly before entering the robot's path. The robot should adapt speed according to visibility and stopping distance rather than assuming that currently unobserved space is empty. Reduced approach velocity provides additional reaction time and can be combined with sensing strategies that improve visibility before committing to the passage.
Prediction-based avoidance helps determine whether a currently open bottleneck will remain open long enough for safe traversal. The planner can compare the robot's estimated entry and exit times with predicted pedestrian or robot occupancy. If another agent is expected to reach the doorway during the crossing interval, the robot may delay entry. This converts doorway avoidance into a space-time scheduling problem rather than a purely spatial collision check.
Robot dynamics must be incorporated because constrained spaces provide little room for late corrections. Heavy AMRs require sufficient distance to decelerate before entering a blocked corridor, while steering-constrained platforms may need additional space for alignment. Maximum acceleration, braking capability, minimum turning radius, control latency, and wheel configuration should therefore influence approach speed and the decision point at which the robot commits to entering.
Once committed to a very narrow passage, excessive replanning can itself become undesirable. Small perception changes may cause repeated steering corrections that produce unstable or uncomfortable motion. A corridor-following or doorway-traversal mode can temporarily emphasize centerline tracking, orientation stability, reduced speed, and controlled clearance while still retaining emergency obstacle detection. Normal local planning can resume after the robot exits the constrained region.
Deadlock recovery is necessary when multiple agents block one another despite preventive coordination. The system should detect prolonged lack of progress and identify which robot has a feasible retreat path. Recovery may require backing toward a previous widening, releasing a reservation, assigning temporary priority, or requesting fleet-level intervention. Random local movements are generally ineffective when the conflict is caused by structural limitations of the environment.
Global planning can reduce bottleneck conflicts before they occur. Routes may be assigned different costs according to corridor width, expected traffic, directionality, or predicted congestion. One-way traffic rules can be introduced for heavily used narrow aisles, while alternative routes can distribute robot flow across the facility. Local avoidance then handles short-term interactions within a traffic structure designed to reduce inherently difficult encounters.
Safety monitoring remains necessary even when priority and reservation logic are used. Communication delays, human entry, localization error, or another robot's failure can invalidate the expected occupancy of a passage. Independent collision monitoring, protective stopping, speed limits, and safety margins should therefore remain active throughout traversal. Coordination improves efficiency, but it should not replace local physical safety mechanisms.
Evaluation should include doorway entry, corridor passing, opposing traffic, pedestrian emergence, blocked exits, localization error, communication loss, and multi-robot deadlock scenarios. Relevant metrics include traversal success, minimum clearance, waiting time, bottleneck throughput, number of stops, retreat frequency, deadlock duration, path smoothness, and safety interventions. Testing should cover both nominal operation and situations where coordination assumptions fail.
Avoidance in narrow corridors and doorways therefore requires more than conventional reactive obstacle avoidance. The problem combines footprint-aware geometry, bottleneck detection, prediction, yielding, priority, reservation, dynamic feasibility, visibility-aware speed control, and deadlock recovery. By recognizing when lateral avoidance is physically impossible, the navigation system can shift from local maneuvering to coordinated space-time management and achieve safer, more efficient passage through constrained environments.
06.09. Dynamic Avoidance Benchmark DWB Crowded Scenes
Dynamic avoidance benchmarking evaluates how reliably and efficiently a mobile robot navigates through environments containing multiple moving agents. In crowded scenes, success cannot be judged only by whether the robot eventually reaches its goal. A meaningful benchmark must measure collision avoidance, clearance, progress, smoothness, interaction quality, computational latency, and recovery from temporary conflicts under repeatable dynamic conditions.
The Dynamic Window Approach and its ROS 2 Navigation2 implementation, commonly associated with the DWB local planner, provide a useful baseline for this evaluation. DWB samples dynamically reachable velocity commands, simulates their short-term trajectories, scores them using multiple critics, and selects a command with favorable overall cost. Its behavior therefore depends strongly on velocity limits, acceleration constraints, simulation horizon, sampling density, footprint, costmaps, and critic configuration.
A crowded-scene benchmark should separate perception, prediction, planning, and control effects whenever possible. If obstacle tracks are inaccurate, an avoidance failure may not indicate a weakness in the local planner itself. Conversely, perfect simulated obstacle states can hide problems that appear with real sensors. Evaluation should therefore include both controlled ground-truth experiments and realistic perception pipelines containing noise, latency, occlusion, and tracking uncertainty.
Scenario design should represent recurring interaction patterns rather than only random moving obstacles. Important cases include head-on encounters, perpendicular crossings, overtaking, merging flows, doorway conflicts, narrow corridors, bidirectional pedestrian streams, stationary groups, and sudden path intrusions. Each scenario exposes different planner characteristics, such as willingness to yield, ability to preserve progress, sensitivity to lateral clearance, and resistance to oscillation.
Crowd density should be varied systematically because planner behavior often changes nonlinearly as free space decreases. Sparse scenes mainly test ordinary dynamic avoidance, while moderate densities reveal interaction quality and trajectory selection. High densities expose local minima, repeated stopping, oscillation, and infeasible velocity choices. Density can be expressed through agents per unit area, average spacing, or the number of interacting agents within the robot's local planning region.
Agent speed and behavioral diversity are equally important. A benchmark containing only constant-speed pedestrians following predictable straight lines may overestimate planner capability. Moving agents should include different walking speeds, accelerations, stopping behavior, direction changes, and yielding tendencies. Some scenarios can remain deterministic for reproducibility, while others use controlled randomization to evaluate robustness across a distribution of interactions.
DWB evaluation should examine how trajectory sampling interacts with crowded geometry. Too few velocity samples can miss useful maneuvers, while very dense sampling increases computation. The simulation time also changes behavior: a short horizon may react late to crossing traffic, whereas a longer horizon may detect conflicts earlier but become constrained by uncertain future obstacle motion. Benchmarking should therefore record both navigation performance and computational cost as these parameters change.
Critic functions are central to DWB behavior because candidate trajectories are ranked according to weighted objectives. Typical considerations include obstacle proximity, path alignment, goal alignment, distance toward the goal, rotation behavior, and oscillation suppression. In crowded scenes, critic weights can substantially alter whether the robot aggressively maintains its route, creates clearance around moving agents, waits for an opening, or repeatedly changes direction.
A benchmark should distinguish collision avoidance from socially acceptable clearance. A robot may technically avoid contact while passing extremely close to pedestrians, which may be unacceptable in human-shared environments. Minimum separation distance, time spent below comfort thresholds, and intrusion into defined personal-space regions provide more informative measurements. These metrics complement binary collision counts and reveal behavior that is geometrically successful but operationally undesirable.
Progress metrics are needed because an extremely conservative robot can obtain a low collision rate simply by stopping. Goal success rate, time to goal, normalized path length, average forward progress, waiting duration, and timeout frequency help distinguish safe navigation from unnecessary paralysis. In dense scenes, successful behavior requires balancing sufficient caution with the ability to exploit temporary openings and continue toward the destination.
Oscillation and control smoothness are particularly important for DWB-style reactive planning. Rapid switching between left and right avoidance choices can produce unstable trajectories even when collisions do not occur. Metrics can include angular velocity variation, acceleration and jerk, command reversals, heading changes, stop-start frequency, and oscillation events. Smooth commands are also important for passenger comfort, payload stability, actuator wear, and predictable interaction with humans.
Dynamic-obstacle interaction should be evaluated in time as well as space. Time-to-collision, time-to-conflict, minimum predicted separation, and duration of conflict exposure can reveal whether the planner reacts early or waits until avoidance becomes urgent. These measures are especially useful when comparing a primarily reactive DWB configuration with prediction-based, velocity-obstacle, ORCA, or learning-based approaches.
Deadlock behavior should be measured explicitly. In narrow or crowded environments, the robot may remain collision-free but fail to make meaningful progress because every sampled trajectory receives an unfavorable score. A benchmark can define a no-progress interval and record deadlock frequency, duration, recovery success, number of replanning attempts, and whether external intervention is required. Such metrics are essential for realistic fleet operation.
Repeatability requires controlled initial conditions and reproducible agent behavior. Robot start and goal poses, obstacle trajectories, timing, map configuration, controller frequency, planner parameters, and random seeds should be recorded. Repeating each scenario across multiple trials allows mean performance to be reported together with variance and tail behavior. Safety-critical evaluation should emphasize worst cases and high-percentile results rather than relying only on averages.
Computational performance must be measured because dynamic avoidance operates under strict real-time constraints. Relevant quantities include controller-cycle time, trajectory evaluation time, CPU or accelerator utilization, missed control deadlines, and end-to-end sensing-to-command latency. P50, P95, P99, and worst-case latency provide a clearer picture than average computation time, particularly when crowded scenes increase the number of obstacles and candidate evaluations.
Parameter sensitivity is another important benchmark dimension. Velocity limits, acceleration bounds, simulation horizon, velocity sampling resolution, inflation radius, costmap update rate, and critic weights can be varied systematically. The objective is not merely to find one configuration that performs well in a single scene, but to determine whether acceptable behavior persists across realistic operating conditions and how strongly performance depends on careful tuning.
Comparative benchmarking should keep sensing, robot dynamics, maps, initial conditions, and scenario scripts consistent when evaluating different avoidance methods. DWB can serve as a classical local-planning baseline against methods such as VO, RVO, ORCA, predictive trajectory optimization, social-force approaches, or learned policies. Differences can then be analyzed in terms of safety, efficiency, smoothness, social behavior, computational demand, and failure modes.
Simulation enables large-scale crowded-scene testing before physical deployment. Thousands of controlled interactions can be executed with variations in crowd density, agent speed, sensor noise, latency, and behavioral uncertainty. However, simulation results should be complemented by physical tests because wheel slip, localization error, perception failures, actuator delay, communication jitter, and real human behavior can produce failure modes that simplified simulation does not reproduce.
A practical benchmark should therefore use a layered progression from deterministic unit scenarios to randomized simulation and finally controlled field trials. Early stages isolate specific behaviors such as crossing or head-on avoidance, while later stages combine multiple interactions and uncertainties. This progression makes failures easier to diagnose and provides stronger evidence that improvements observed in simulation remain meaningful on the physical robot.
Dynamic avoidance benchmarking in crowded scenes ultimately measures whether a navigation system remains safe, progressive, stable, and computationally predictable as interaction complexity increases. DWB provides a valuable reference because its sampled trajectories and critic-based decisions are interpretable and tunable. A rigorous benchmark extends beyond collision rate to expose clearance, deadlock, oscillation, latency, robustness, and worst-case behavior, creating a foundation for comparing increasingly advanced dynamic avoidance methods.
06.10. Human Aware Robot Navigation Social Compliance Case
Human-aware robot navigation extends conventional collision avoidance by treating people as social participants rather than moving obstacles defined only by geometry. A socially compliant robot must remain physically safe while moving in ways that humans can easily interpret and accept. This requires reasoning about personal space, pedestrian flow, right-of-way, visibility, comfort, predictability, and culturally or contextually appropriate interaction.
Physical collision avoidance represents only the minimum requirement. A robot can maintain a technically collision-free trajectory while still making people uncomfortable by passing too closely, approaching rapidly from behind, cutting between interacting people, or forcing pedestrians to change direction abruptly. Human-aware navigation therefore introduces social constraints and costs that shape robot motion before a geometric collision becomes imminent.
Personal space is commonly represented as a spatial region surrounding each person. Unlike a fixed circular safety radius, this region may be asymmetric and dependent on motion. A moving pedestrian often requires greater clearance in front than behind or to the side because the forward region reflects likely future movement and perceptual attention. Navigation costs can therefore vary according to relative position, orientation, speed, and interaction context.
Proxemic models provide a useful framework for representing socially meaningful distance. Different spatial zones can indicate intimate, personal, social, or public interaction ranges, although appropriate distances vary with environment and culture. For navigation, these concepts are usually converted into continuous cost fields rather than rigid boundaries, allowing the planner to trade path efficiency against increasing discomfort as the robot approaches a person.
Pedestrian direction should influence avoidance behavior. When encountering a person head-on, the robot should select a clear avoidance side early enough to make its intention understandable. During overtaking, it should maintain additional lateral clearance and avoid sudden acceleration near the pedestrian. When crossing a pedestrian flow, the robot should seek a suitable gap instead of forcing its way through a sequence of closely spaced people.
Predictability is an important component of social compliance. Humans estimate robot intent from velocity, heading, path curvature, and changes in motion. Frequent left-right oscillation or abrupt stopping makes the robot difficult to interpret even if every individual command is collision-free. Smooth trajectories, early directional commitment, moderate acceleration, and consistent yielding behavior make robot motion easier for nearby people to anticipate.
Visibility also affects socially appropriate navigation. A robot approaching from a person's front is generally easier to notice than one approaching quickly from behind or emerging from an occluded corner. Human-aware planning can penalize trajectories through low-visibility interaction regions and reduce speed near doorways, blind corners, shelves, and intersections. The robot should preserve enough stopping distance for people who may appear unexpectedly.
Groups of people require reasoning beyond independent personal-space models. Two people facing each other may be engaged in conversation, and walking groups may form a socially meaningful unit. Passing directly between group members can be disruptive even when sufficient geometric space exists. Group-aware navigation can detect formations and assign additional cost to socially occupied regions, encouraging the robot to pass around the group when feasible.
Queues are another important social structure. A navigation system that sees only free geometric space may attempt to cross between people waiting in line or stop inside the queue. Semantic interpretation can identify queue direction, service points, and expected pedestrian behavior. The robot can then route around the queue or cross it only at locations where disruption is minimized and sufficient clearance is available.
Doorways, elevators, and narrow corridors create implicit right-of-way interactions. A robot approaching a doorway while a pedestrian is exiting should normally yield rather than entering simultaneously. At elevators, people leaving the cabin are commonly allowed to exit before others enter. Encoding such contextual conventions allows the robot to behave in ways that align with human expectations instead of relying exclusively on shortest-path optimization.
Social compliance also involves appropriate speed. High speed may be acceptable in an empty industrial corridor but uncomfortable in a dense human-shared area. Speed can be adapted according to pedestrian density, minimum separation, visibility, environmental semantics, and predicted interaction risk. Slowing before close encounters provides additional reaction time and communicates cautious intent, but excessive stopping should be avoided because it can make the robot obstructive.
Trajectory prediction strengthens human-aware navigation by estimating where pedestrians are likely to move rather than reacting only to their current locations. Historical motion, body orientation, scene structure, and surrounding agents can contribute to future trajectory estimates. Multimodal prediction is valuable because a pedestrian near an intersection may continue straight, turn, stop, or interact with another person, creating several plausible future behaviors.
Interaction-aware prediction recognizes that pedestrians may react to the robot itself. A trajectory forecast that assumes humans continue unchanged regardless of robot motion can become inaccurate during close encounters. More advanced models represent reciprocal influence, allowing the planner to estimate how a person may yield, accelerate, change direction, or preserve personal space in response to the robot's planned trajectory.
Social navigation costs can be integrated into conventional local or trajectory optimization planners. Candidate trajectories may be evaluated according to obstacle clearance, goal progress, path alignment, personal-space intrusion, pedestrian disturbance, visibility, group disruption, and control smoothness. The planner can then select motion that remains efficient while avoiding trajectories that are physically safe but socially undesirable.
Learning-based approaches can capture interaction patterns that are difficult to encode manually. Imitation learning can reproduce socially appropriate demonstrations, while reinforcement learning can optimize policies using rewards related to safety, progress, comfort, and pedestrian disturbance. Graph neural networks and attention-based models can represent relationships among multiple people, groups, and the robot, supporting reasoning in crowded and dynamically changing environments.
Learned social behavior nevertheless requires careful safety separation. A policy trained from human demonstrations or simulated crowds may behave unexpectedly in rare configurations, unfamiliar cultures, or environments outside its training distribution. Deterministic collision monitoring, speed limits, emergency stopping, or Control Barrier Function safety layers can therefore remain independent of the learned social-navigation component.
A practical human-aware architecture can use layered decision making. Perception detects and tracks people, semantic reasoning identifies groups and contextual regions, prediction estimates future motion, and a social planner generates a preferred trajectory. A local controller executes this trajectory while an independent safety layer monitors minimum separation and braking requirements. Continuous replanning updates behavior as pedestrian intentions become clearer.
Evaluation should measure both physical safety and human impact. Relevant metrics include collision rate, minimum separation, personal-space intrusion, pedestrian path deviation, pedestrian speed change, robot waiting time, navigation efficiency, trajectory smoothness, and unnecessary stops. Human-subject studies can additionally measure perceived safety, comfort, predictability, trust, and whether participants believe the robot behaved appropriately in shared space.
Scenario-based testing should include head-on encounters, overtaking, perpendicular crossings, groups, queues, doorway interactions, blind corners, elevator entry, narrow corridors, and dense bidirectional pedestrian flows. Variations in pedestrian speed, density, attention, and willingness to yield reveal whether behavior remains stable across realistic interactions. Rare cases such as sudden stopping or unexpected path changes are particularly important for safety validation.
Human-aware robot navigation ultimately requires the robot to optimize not only where it can move, but how its movement affects nearby people. Social compliance emerges from combining physical safety, personal-space modeling, prediction, contextual conventions, group awareness, smooth control, and understandable intent. A robot that navigates according to these principles can share human environments more safely and naturally while maintaining efficient progress toward its operational goals.
Author: Richard Minsoo Go
LinkedIn: Richard Minsoo Go










Top comments (0)