Collision Avoidance System (CAS)

This subsystem is responsible for creating a real-time, unified model of the environment. It ingests raw data from heterogeneous sensors and fuses it into a coherent world state.

• Primary Sensors: LiDAR for precise 3D point clouds, radar for velocity and long-range detection in poor weather, and cameras for semantic classification.
• Sensor Fusion Algorithm: Combines data probabilistically (e.g., using a Kalman Filter or Bayesian network) to overcome individual sensor limitations (e.g., camera blindness in low light, LiDAR sensitivity to rain).
• Output: A continuously updated Occupancy Grid or list of tracked dynamic objects with attributes like position, velocity, and classification.

This component forecasts the future motion of all dynamic obstacles to enable proactive, rather than purely reactive, avoidance. It transforms a snapshot of the world into a time-series prediction.

• Models Used: Simple constant-velocity kinematic models for short-term prediction, or more complex machine learning models (e.g., LSTMs, Graph Neural Networks) trained on trajectory data to predict pedestrian or vehicle intent.
• Critical Output: For each tracked object, it generates multiple probable future paths, often with associated probabilities. This feeds directly into the Collision Risk Assessment module.
• Example: Predicting if a pedestrian at a curb will step into the roadway or remain stationary.

This is the decision-making core that quantifies the danger posed by each predicted interaction. It translates geometry and motion into actionable risk metrics.

• Key Metrics: Calculates Time to Collision (TTC), Distance to Closest Point of Approach (DCPA), and constructs Collision Cones to determine if the current course is unsafe.
• Probabilistic Framework: Often uses a Bayesian approach to combine sensor uncertainty, prediction confidence, and object type (e.g., a child vs. a static cone) into a single risk score.
• Thresholding: Compares calculated risks against predefined Safety Margins to trigger alerts or invoke the planner.

This subsystem generates a kinematically and dynamically feasible, collision-free trajectory and translates it into low-level actuator commands.

• Planner Types: May use reactive methods like the Dynamic Window Approach (DWA) for immediate obstacles, or optimization-based Model Predictive Control (MPC) for smooth, longer-horizon planning that respects vehicle dynamics.
• Constraint Integration: The planner's optimization cost function includes terms for comfort, efficiency, and progress toward the goal, while hard constraints enforce collision avoidance (often via Control Barrier Functions) and physical limits.
• Output: A sequence of steering angles and acceleration/deceleration commands sent to the drive-by-wire system.

The real-time software and hardware foundation that guarantees deterministic execution and enforces fail-safe behavior. This is critical for functional safety certification (e.g., ISO 26262, IEC 61508).

• Real-Time Operating System (RTOS): Ensures task scheduling meets strict deadlines. The Worst-Case Execution Time (WCET) of the CAS loop must be characterized and guaranteed.
• Runtime Assurance (RTA): A simpler, formally verified safety monitor runs in parallel to the primary planner. It can override commands with a verified safe action (like an Emergency Stop) if a violation is predicted.
• Diagnostics: Continuous monitoring for sensor degradation, software faults, and hardware failures.

A cooperative component that extends perception beyond line-of-sight by exchanging intent and state information with other agents and infrastructure.

• Protocols: Uses Dedicated Short-Range Communications (DSRC) or Cellular V2X (C-V2X) to broadcast Basic Safety Messages (BSMs).
• Information Shared: Includes precise GNSS position, speed, heading, acceleration, and vehicle size. Can also share planned trajectory or intent (e.g., 'turning left at next intersection').
• Use Case: Enables avoidance of collisions at blind intersections or with vehicles obscured by large obstacles, forming a cooperative perception field.

A CAS relies on sensor fusion from LiDAR, radar, cameras, and ultrasonic sensors to build a coherent world model. Critical challenges include:

• Time synchronization and spatial calibration (extrinsic/intrinsic) between heterogeneous sensors.
• Handling sensor dropout and conflicting data (e.g., camera failure in low light).
• Propagating measurement uncertainty through the perception stack to the planning layer, often using probabilistic frameworks like Kalman Filters or Particle Filters. Failure to quantify uncertainty can lead to overconfident and unsafe maneuvers.

Collision avoidance is a hard real-time problem. Missing a deadline can mean a collision. Implementation requires:

• Deterministic algorithms with bounded computation. Algorithms like ORCA and MPC must be analyzed for their Worst-Case Execution Time (WCET).
• Efficient collision checking using data structures like Axis-Aligned Bounding Box (AABB) trees or occupancy grid lookups.
• Priority-aware scheduling within the robot's middleware to ensure the CAS loop preempts less critical processes. Certification for safety-critical domains (e.g., ISO 26262 for vehicles) demands rigorous WCET verification.

To account for system latencies, tracking errors, and physical dynamics, safety margins are non-negotiable. Standard practices include:

• Obstacle Inflation: Geometrically expanding all perceived obstacles by a radius that incorporates the agent's footprint, stopping distance, and a confidence buffer.
• Dynamic margins that scale with velocity; higher speeds require larger temporal buffers.
• Multi-layer margins: A smaller, "hard" margin for emergency stops and a larger, "soft" margin for comfortable path planning. This creates a gradient of risk, allowing the system to escalate responses.

A CAS must remain safe even when its primary planning algorithm fails. This is achieved through defense-in-depth architectures:

• Emergency Stop (E-Stop) Protocol: A dedicated, high-integrity channel for triggering a controlled halt upon detecting a critical failure or unavoidable collision.
• Runtime Assurance (RTA): A simpler, formally verified safety monitor runs in parallel with the main CAS. It checks the proposed action against a Control Barrier Function (CBF) or similar constraint. If a violation is predicted, the RTA overrides with a safe fallback command.
• Graceful degradation: Falling back from cooperative (RVO) to non-cooperative (VO) to purely reactive (APF) strategies as communication or sensing degrades.

In shared spaces, decentralized avoidance can lead to livelock or deadlock (e.g., two robots facing each other in a narrow corridor). Mitigation strategies include:

• Reciprocal algorithms like RVO/ORCA that guarantee convergence under ideal conditions.
• Higher-level arbitration from the fleet orchestrator to assign temporary priorities or modify goals.
• Intent signaling via V2X communication or explicit light signals to break symmetry.
• Deadlock detection algorithms that identify cyclic waits and trigger a centralized resolution protocol, often part of the Exception Handling Framework.

Testing a CAS solely in the real world is impractical and dangerous. A rigorous V&V pipeline is essential:

• Simulation: High-fidelity Sim-to-Real environments (e.g., NVIDIA Isaac Sim) to stress-test against thousands of edge cases, including sensor noise and adversarial agent behavior.
• Formal verification: For core algorithms, using methods to prove liveness (a goal is eventually reached) and safety (collisions never occur) under defined assumptions.
• Scenario-based testing: Executing against standardized test suites (e.g., NHTSA scenarios for vehicles) and fault injection to test failure mode responses.
• Operational Design Domain (ODD): Explicitly defining the environmental, geographic, and traffic conditions where the CAS is validated to perform safely.

The Velocity Obstacle (VO) is a geometric algorithm for collision checking. For a moving agent, it defines the set of all velocities that would cause a collision with another moving obstacle within a specified time horizon. The agent simply selects any velocity outside this forbidden set.

• Core Concept: Projects obstacle velocities into the agent's velocity space.
• Application: Foundational for many reactive, decentralized multi-agent algorithms.

Optimal Reciprocal Collision Avoidance (ORCA) is a formal, velocity-based algorithm for cooperative multi-agent navigation. It extends the Reciprocal Velocity Obstacle (RVO) concept by efficiently computing a half-plane of permitted velocities for each agent, assuming all share responsibility for avoidance.

• Key Feature: Provides provably collision-free and oscillation-free trajectories.
• Use Case: Standard for dense, decentralized navigation in robotics simulations and warehouses.

Time to Collision (TTC) is a fundamental risk metric. It estimates the time remaining before two objects on a constant relative velocity course will collide if no evasive action is taken.

• Calculation: Derived from relative distance and relative speed.
• Critical Threshold: A TTC below 2-3 seconds often triggers emergency protocols in automotive and robotic systems.

Model Predictive Control (MPC) for collision avoidance is an optimization-based strategy. It repeatedly solves a finite-horizon optimal control problem to compute a sequence of control inputs that satisfies the agent's dynamic constraints while avoiding predicted collisions.

• Advantage: Explicitly handles constraints and optimizes for smoothness and energy efficiency.
• Trade-off: Computationally intensive, requiring powerful onboard processors.

A Predictive Control Barrier Function (CBF) is a mathematical tool for safety-critical control. It formally guarantees a system's state will remain within a predefined safe set (e.g., a collision-free region) over a future time horizon by acting as a filter on the control input.

• Formal Guarantee: Provides rigorous safety proofs, crucial for certification.
• Integration: Often used in conjunction with MPC or learning-based controllers as a safety layer.

Runtime Assurance (RTA) is a safety architecture for complex autonomous systems. It employs a simple, verified safety monitor or safety controller that can override the primary, potentially complex controller (e.g., a neural network) if its actions are predicted to violate a safety property like collision avoidance.

• Architecture: Enables the use of high-performance, less verifiable algorithms while maintaining a safety guarantee.
• Example: An RTA layer could enforce a minimum distance to obstacles regardless of the primary planner's output.