Obstacle Avoidance¶

Continuously keep a moving robot or autonomous vehicle out of occupied and predicted-danger regions while preserving a feasible route toward its task goal under sensing, dynamics, and reaction-time limits.

Version: v3 · 2026-09-06 · History
Domain-specific #: 2411
Origin domain: engineering
Subdomain: robotics
Aliases: Robot obstacle avoidance, Collision avoidance in robotics

Core Idea¶

Obstacle Avoidance is the robotics capability that converts evidence about occupied or imminently dangerous space into motion that remains collision-free while still advancing an assigned task. A complete instance binds a moving body, sensed or modeled obstacles, the body's geometry and dynamics, a safety margin, a goal or progress criterion, and a decision process that selects an admissible motion before the available reaction horizon closes. Merely detecting an object is insufficient; the system must alter or constrain motion. Merely stopping forever is also not the ordinary identity, because avoidance normally preserves feasible progress whenever such progress exists.

In configuration-space planning, the robot's body is represented by a configuration \(q\in\mathcal C\). Obstacles induce a forbidden region \(\mathcal C_{obs}\), and collision-free configurations form \(\mathcal C_{free}=\mathcal C\setminus\mathcal C_{obs}\). A global motion-planning problem seeks a continuous path \(\tau:[0,1]\to\mathcal C_{free}\) from an initial configuration to a goal configuration. LaValle's treatment shows how this geometric conversion separates the body's shape from the path search and how collision checking, sampling, roadmaps, and differential constraints enter the planning problem.^[1] Obstacle avoidance is broader than one global planner: it also covers feedback policies that revise velocity or control as observations change.

Khatib's artificial-potential-field method is a canonical reactive architecture. Obstacles exert repulsive influence while the goal exerts attractive influence, and the resulting field contributes directly to operational-space control. Its importance is structural rather than prescriptive: obstacle response can be distributed into lower control levels instead of waiting for a complete high-level replan, and a time-varying field can represent moving obstacles.^[2] Potential fields can create local minima, oscillation, and narrow-passage failures, so the named abstraction cannot be reduced to this one technique.

The dynamic-window approach demonstrates another architecture. It searches reachable translational and rotational velocities over a short horizon, excludes commands that cannot stop before collision, and optimizes among the remaining commands for progress, clearance, and speed. Because the window is derived from the platform's acceleration limits, it ties avoidance to executable dynamics rather than to a geometric path alone.^[3] Global and local layers are often combined: a route supplies long-range intent while a reactive controller handles newly sensed obstacles. The invariant is the perception-to-admissible-motion loop under a safety constraint, not a specific sensor, map, cost function, or algorithm.

Structural Signature¶

Moving agent. A robot, manipulator, vehicle, or other controlled body has nonzero spatial extent and a current state.
Task intent. A destination, route, tracking target, or productive motion distinguishes avoidance from indefinite immobilization.
Obstacle model. Static or moving objects induce occupied or predicted-danger regions in workspace or configuration space.
Perception channel. Sensors, a map, communicated tracks, or prior geometry supply obstacle evidence with uncertainty and latency.
Geometric envelope. Robot shape, swept volume, localization error, and clearance margin determine which states are unsafe.
Dynamic feasibility. Velocity, acceleration, turning radius, braking distance, and actuator limits restrict executable evasive motions.
Prediction horizon. Relative motion is projected far enough ahead to detect a collision before reaction becomes impossible.
Admissibility test. Candidate paths or controls are rejected when their swept motion intersects forbidden space or violates stopping safety.
Selection rule. The system ranks safe candidates by progress, clearance, smoothness, effort, or another declared objective.
Feedback update. New observations revise the state, obstacle estimate, and selected control before stale plans become dangerous.
Failure response. No-safe-motion, sensor-loss, and localization-loss states trigger stopping, retreat, replanning, or escalation.
Safety–progress invariant. Collision exclusion has priority while feasible goal progress is retained rather than silently discarded.

What It Is Not¶

Not obstacle detection. Detection identifies an object; avoidance converts that evidence into a safe motion decision.
Not collision detection alone. A geometric intersection query is one test inside the control loop, not the complete capability.
Not all path planning. Planning can occur in an obstacle-free model or optimize objectives unrelated to collision risk.
Not SLAM. Simultaneous localization and mapping estimates pose and environment; avoidance may consume those estimates without building them.
Not an emergency stop. Stopping is an allowed terminal response, but ordinary avoidance seeks an alternative that preserves progress.
Not one algorithm. Potential fields, velocity obstacles, dynamic windows, roadmaps, sampling planners, and learned policies are implementations.
Not guaranteed safety from nominal testing. Sensing gaps, uncertainty, latency, dynamics, and model error must be included in the safety case.
Not navigation in full. Localization, route selection, mission logic, and destination choice extend beyond the avoidance subproblem.

Scope of Application¶

Obstacle Avoidance is literal when a controlled moving body must respond to spatial obstruction while continuing a goal-directed navigation task under finite sensing and actuation horizons.

Mobile robots. Indoor service and warehouse robots combine route guidance with short-horizon local collision avoidance.
Robot manipulators. Link geometry and swept volumes constrain joint or operational-space motion around fixtures, people, and workpieces.
Autonomous road vehicles. Predicted trajectories, road constraints, and emergency envelopes shape evasive steering and braking.
Aerial and marine vehicles. Three-dimensional dynamics and environmental disturbances alter reachable safe maneuvers.
Human–robot collaboration. Conservative uncertainty bounds and stop responses protect people whose motion is only partially predictable.
Multi-robot systems. Other agents may be treated as dynamic obstacles, although reciprocal coordination requires additional protocol assumptions.
Assistive mobility. Powered chairs and guidance systems use avoidance while preserving operator intent and accessible clearance.
Simulation and verification. Scenario suites test minimum separation, reaction latency, deadlock, and recovery before deployment.

Clarity¶

State the robot geometry, state variables, obstacle representation, sensor horizon, localization and perception uncertainty, control cycle, safety margin, dynamics, prediction horizon, and progress objective. Distinguish static from dynamic obstacles and geometric collision-free paths from dynamically executable trajectories. Specify whether the controller is complete, probabilistically complete, resolution complete, or only heuristic. Report the behavior when no safe command exists. If using a learned component, separate learned perception or scoring from the independently enforced safety constraint. Avoid claiming formal collision freedom when the proof excludes sensor error, unmodeled actors, delay, or actuator saturation. State whether other agents cooperate, remain passive, or act adversarially.

Manages Complexity¶

The raw problem couples high-dimensional geometry, uncertain perception, prediction, discrete route choices, continuous dynamics, and millisecond control. Obstacle-avoidance architectures manage it by converting bodies to forbidden configuration regions, separating long-range route intent from short-range feedback, pruning unsafe controls before optimization, and repeatedly solving a bounded local decision. This decomposition makes operation tractable, but it introduces interfaces where stale maps, inconsistent frames, horizon effects, and local minima can cause failure. Increasing clearance can destroy passage feasibility; increasing horizon can exceed computation time; optimizing smoothness can delay escape. A robust design therefore preserves explicit safety invariants, exposes uncertainty, and treats no-solution states as outcomes rather than hiding them behind a cost minimum.

Abstract Reasoning¶

Declare the controlled body's geometry, current state, task goal, and actuation limits.
Transform sensed or mapped obstacles into occupied and buffered danger regions.
Propagate state and obstacle uncertainty over the chosen prediction horizon.
Generate candidate paths, velocities, or controls that are dynamically reachable.
Reject candidates whose swept volumes violate the collision or braking constraint.
Rank remaining candidates by declared progress, clearance, smoothness, and effort criteria.
Execute only the bounded prefix for which observation and model assumptions remain current.
Update perception and localization, then repeat the admissibility calculation.
Detect deadlock, oscillation, local minima, and empty safe-control sets explicitly.
Invoke stop, retreat, global replanning, or human escalation under the declared failure policy.

Knowledge Transfer¶

The strict parent is Navigation. Obstacle Avoidance presupposes a positioned moving agent and a goal-directed motion task; its residual is the continual exclusion of collision-producing paths or controls under obstacle uncertainty and finite dynamics. Navigation transfers localization, direction, route, and progress concepts, while the domain node adds body geometry, swept-volume safety, dynamic feasibility, and reactive replanning. Algorithm is an internal implementation neighbor, but the capability can be realized by many algorithms and is therefore not one algorithmic subtype.

Examples¶

Canonical¶

A disk-shaped indoor robot has a lidar map, pose covariance, speed and acceleration bounds, and a target waypoint. The controller inflates each obstacle by the robot radius plus an uncertainty margin. Every cycle it samples velocities reachable during the next interval, rejects any velocity whose projected arc cannot stop before entering the inflated obstacle set, and selects a remaining velocity balancing heading, clearance, and speed. When a person steps into the route, the admissible window changes immediately; the robot slows and turns, then resumes the route when a safe corridor reopens.^[3]

Mapped back: localized moving body + inflated obstacle regions + reachable velocity set → unsafe-command pruning → safe progress with feedback.

Applied / In Practice¶

A manipulator carries a long workpiece through a cell. Checking only the gripper point would miss collisions by the links and load. The planner models the full configuration-space obstacle region, finds a global corridor, and executes it with an operational-space repulsive field that reacts to a newly placed fixture.^[2] If the local field stalls in a minimum, the supervisor returns to the global planner rather than treating zero velocity as successful completion.

Mapped back: full-body swept geometry + planned corridor + reactive field → collision exclusion → detected stall → supervised replanning.

Structural Tensions¶

Safety vs. progress. Larger margins reduce collision risk but can erase every corridor. Diagnostic: Which uncertainties justify the margin, and what happens when the safe set is empty?
Global completeness vs. reactive speed. Rich search finds routes but may miss the control deadline. Diagnostic: Which decision is delegated to each time scale?
Prediction vs. model error. Longer forecasts anticipate motion but amplify behavioral uncertainty. Diagnostic: At what horizon does added prediction cease improving safe decisions?
Smoothness vs. escape. Penalizing control change can retain a trajectory toward danger or deadlock. Diagnostic: Can the objective override the hard admissibility filter?
Individual avoidance vs. reciprocal coordination. Treating agents as obstacles may create freezing or oscillation. Diagnostic: Which agents share conventions or intent?
Autonomous residual vs. generic Navigation. Every navigation system encounters constraints. Diagnostic: Does the system explicitly estimate obstacles and alter motion to maintain collision exclusion?

Structural–Framed Character¶

A moving body, task intent, obstacle evidence, geometric safety envelope, dynamic feasibility, admissibility test, feedback update, and no-safe-motion response are structural. Sensor type, map representation, planner, cost weights, platform, environment, speed, and communication are framed. The abstraction guarantees neither collision immunity nor mission completion without assumptions; it organizes how those claims must be made and tested.

Structural Core vs. Domain Accent¶

The transferable skeleton is goal-directed Navigation under constraints. The robotics accent is configuration-space obstruction, swept volume, sensing uncertainty, stopping distance, reachable controls, collision prediction, local reaction, and recovery from deadlock. Removing the goal yields safety monitoring or stopping; removing obstacle-triggered motion revision yields ordinary navigation or planning.

Navigation is the strict parent by composition/presupposition: avoidance retains a destination-directed moving agent and route progress while adding an obstacle-conditioned safety loop. The edge does not make every navigation system an obstacle avoider and does not collapse avoidance into localization, mapping, or a particular Algorithm.

The prospective workspace queue contains one strict upward edge to prime:navigation. No live DAG mutation is authorized.

Relationships to Other Abstractions¶

Current abstraction Obstacle Avoidance Domain-specific

Parents (1) — more general patterns this builds on

Obstacle Avoidance is a kind of Navigation Prime

Navigation is the strict parent by composition/presupposition: avoidance retains a destination-directed moving agent and route progress while adding an obstacle-conditioned safety loop.

Hierarchy path (1) — routes to 1 parentless root

Obstacle Avoidance → Navigation

Neighborhood in Abstraction Space¶

Obstacle Avoidance sits in a sparse region of the domain-specific corpus (94^th percentile for distinctiveness): few abstractions share its structure, so a faithful description tends to retrieve it precisely.

Family — Unclustered & Miscellaneous (1565 abstractions)

Nearest neighbors

Computed from structural-signature embeddings · 2026-09-08

Not to Be Confused With¶

Collision Detection. Tests whether two geometric objects or swept volumes intersect.
Path Planning. Searches for a route and may be global, offline, or unaware of new obstacles.
Motion Planning. Adds kinematics and dynamics but includes tasks beyond obstacle response.
SLAM. Jointly estimates pose and a map from observations.
Emergency Braking. Selects deceleration rather than a general evasive motion.
Artificial Potential Field. One reactive implementation with known local-minimum behavior.
Dynamic Window Approach. One velocity-space local avoidance implementation.

References¶

[1] Steven M. LaValle, Planning Algorithms (Cambridge University Press, 2006), especially chs. 4–8, https://lavalle.pl/planning/. registry ↩

[2] Oussama Khatib, “Real-Time Obstacle Avoidance for Manipulators and Mobile Robots,” International Journal of Robotics Research 5, no. 1 (1986): 90–98, https://doi.org/10.1177/027836498600500106. registry ↩a ↩b

[3] Dieter Fox, Wolfram Burgard, and Sebastian Thrun, “The Dynamic Window Approach to Collision Avoidance,” IEEE Robotics & Automation Magazine 4, no. 1 (1997): 23–33, https://doi.org/10.1109/100.580977. registry ↩a ↩b