On-Demand Collision Avoidance Techniques

• On-demand collision avoidance is a system that dynamically triggers reactive maneuvers using local sensors and rapid decision-making.
• It employs geometric rules, optimization methods, and learning-based strategies to enforce safety constraints and minimize collision risks.
• Decentralized coordination allows scalable application across autonomous vehicles, drones, marine vessels, and robots in diverse environments.

On-demand collision avoidance encompasses a spectrum of algorithms, architectures, and control paradigms that enable autonomous agents—spacecraft, ground vehicles, marine vessels, aerial drones, and mobile robots—to detect imminent threats and execute maneuvers that maintain safety in real time, often without prior global planning or centralized oversight. The defining characteristic is the dynamic, context-sensitive activation: avoidance behavior is triggered immediately upon risk detection, leveraging onboard sensors, local model predictions, or inter-agent communications, and is resolved through a combination of optimization, learning, reactive control, or geometric rule sets.

1. Foundational Principles and Mathematical Formalism

On-demand collision avoidance systems are unified by several core principles:

• Reactive triggering: Maneuvering decisions are initiated strictly in response to detected threats, often using minimal prediction or local sensing.
• Constraint enforcement: Safety is strictly enforced via state-space constraints, reachable set separation, barrier functions, or explicit trajectory exclusions.
• Decentralization and scalability: Architectures favor per-agent computations, enabling operation in large fleets, swarms, or dense traffic without centralized coordination.
• Real-time performance: Algorithmic latency is sufficiently low (typically <50 ms) to guarantee timely actuation in rapidly evolving dynamic environments.

Formally, approaches span from direct geometric rules (minimum separation, closest-point-of-approach) to fully optimized control (convex QP, nonlinear MPC), and advanced stochastic formulations (POMDP, DRQN-based policies (Bourriez et al., 2023)). For multi-agent systems, problem statements are often formulated as constrained optimization over predicted state trajectories:

$$\min_{u}\ J(x, u)\quad \text{subject to}\ \mathrm{collision\_constraints}(x, u) \geq r_{\min}$$

where $J$ encodes dynamic performance (fuel, smoothness, deviation), and $\mathrm{collision\_constraints}$ maintain required separation or safety margins, which may be adaptively inserted only when needed (Gupta et al., 2023).

2. Algorithms and Control Architectures

2.1 Geometric and Rule-based Approaches

• CPA (Closest Point of Approach) methods (Pham et al., 2015): Direct calculation of intersection and time-to-collision, employing simple heading or speed adjustments when a risk threshold is exceeded. Suitable for UAVs and surface vessels with low computational resources.
• COLREGs-compliant cost functionals in marine settings: For autonomous vessels (ASVs), soft cost penalties incorporating international maritime safety codes yield provably legal maneuvers in head-on, crossing, and overtaking scenarios (Eriksen et al., 2019).

2.2 Model Predictive Control (MPC) and Optimization

• Branching-course MPC (BC-MPC): Enumerates a tree of speed/course trajectories, attaches soft penalties for static/moving obstacles, and selects the least-cost branch, executing only the first maneuver segment (Eriksen et al., 2019).
• Convex and non-convex constraint embedding: In on-orbit robotics, collision avoidance is enforced via differentiable KKT conditions derived from convex distance subproblems embedded into the overall optimal control program (Tavana et al., 11 Apr 2024).
• Variable horizon MPC with versatile on-demand constraint activation: Collision constraints are linearly inserted only for the earliest predicted conflict index, minimizing QP dimension at each robot/time step and facilitating adaptive horizon control via RL (Gupta et al., 2023).

2.3 Learning-based and Hybrid Methods

• Deep RL/POMDP for space traffic management: Onboard agents use DRQN architectures to process noisy, partial orbital observations, autonomously planning impulsive maneuvers to minimize risk and fuel cost (Bourriez et al., 2023).
• Learning-based conflict resolution for UAS swarms: L2F/LNF utilizes supervised LSTM classifiers for fast separation decisions, followed by pairwise convex MPC, yielding millisecond-level online collision resolution in large, mission-constrained airspaces (Rodionova et al., 2021, Rodionova et al., 2020).
• SAFER RL+trajectory search: Real-world robot learning is stabilized by hard emergency-brake triggers, and RL policy restricts dynamic window search to efficient local corrections, yielding safety with minimal intervention (Srouji et al., 2022).

3. Sensor Modalities, State Estimation, and Perception Pipelines

Collision avoidance performance is tightly coupled to sensing and state estimation:

• ADS-B, Electronic Conspicuity, and Cooperative positions: Enables geometric conflict detection for air and drone platforms; DACM employs fused EC feeds and Kalman-smoothing for cm-level position accuracy and fast risk scoring (Kuru et al., 2023).
• Vision-based approaches: Sequential Spatial Network (SSN) fuses stacked top-down camera frames with convolutional and attention modules to predict safe trajectories in dense urban driving scenarios (Li et al., 2023). Biologically inspired control laws rely solely on monocular estimation of bearing rate, loom (TTC), and line-of-sight angle (Marinho et al., 2021).
• Depth and TTC refinement with machine learning: Agile quadrotors process RGB-D data with monocular depth completion, polynomial fitting, and per-pixel TTC estimation, dynamically constructing control barrier functions for reactive NMPC (Saviolo et al., 18 Sep 2024).
• Sensor-data preprocessing and calibration: Camera feeds processed through homography and neural networks enable precise distance estimation for highway scenarios, followed by Kalman filtering for robust velocity/acceleration inference (Chen et al., 26 Apr 2025).

4. Multi-Agent Coordination and Decentralized Protocols

Advanced collision avoidance incorporates decentralized cooperation:

• Desired vs. planned trajectory broadcasting: Vehicles compute reference-free desired paths (ignoring others), planned paths (considering others), and weakly penalize overlap, yielding emergent cooperative behavior without leader election (Wartnaby et al., 2019).
• Pairwise robust separation via convex optimization: UAS and UAVs jointly shrink admissible control sets (ellipsoidal constraints) so that reachable tubes remain disjoint up to a safety horizon, all solved onboard via fast SDP (Zhou et al., 2015).
• Priority ordering and tube-shrinking: For $N$-agent urban airspaces, systematic pairwise scheduling and conflict tube shrinking resolve live-lock, ensuring finite convergence and bounded safety guarantees (Rodionova et al., 2021).

5. Emergency Integration and Robustness to Edge Cases

Effective frameworks feature multi-level decision logic and adaptive responses:

• Emergency steering and braking strategies: Road vehicle frameworks combine safe-distance models, dynamic collision hazard detection, MPC-based lateral planning, and longitudinal deceleration, switching modes as threat proximity evolves (Chen et al., 2023).
• Torque vectoring stabilizers and hierarchy: For structured environments, torque allocation at the wheel level ensures yaw and lateral stability when sharp maneuvers are required, maintaining performance even under low friction and high curvature (Taherian et al., 2020).
• Integrated reward shaping and RL training: High-risk autonomous driving leverages MDPs incorporating both leading and trailing vehicle behaviors, training policies for adaptive cruising and emergency braking, outperforming classical AEB criteria (Chen et al., 26 Apr 2025).

6. Empirical Validation, Performance Metrics, and Scalability

Experiments and field trials demonstrate viability across platforms and environments:

• Spacecraft collision risk reduction: Onboard DRQN policies halve post-maneuver $P_c$ versus threshold-trigger baselines, with 20 ms inference latency and persistent fuel savings (Bourriez et al., 2023).
• ASV sea trials: BC-MPC maintains 130–170 m minimum separation in head-on and overtaking scenarios, complying with all COLREGs and producing responsive, low-wobble trajectories (Eriksen et al., 2019).
• Urban airspace density scaling: LNF resolves 91–99.9% of pairwise UAS conflicts at up to 70 agents, maintaining STL mission guarantees, with convex MPC solves in 15–200 ms per step (Rodionova et al., 2021).
• Robot navigation efficiency: SAFER RL+search runs full avoidance cycles in 20 ms, with collision rates near zero for tight indoor passages compared to pure planning or RL methods (Srouji et al., 2022).
• Reactive quadrotor NMPC: Unified TTC/CBF-based NMPC achieves 2 ms solver time, robust collision avoidance, and high-speed flight in obstacle-rich environments, with fully onboard computation (Saviolo et al., 18 Sep 2024).

7. Limitations, Extensions, and Future Directions

Current limitations and research directions include:

• Action space discretization in RL approaches (e.g., need for continuous-action actor–critic in spacecraft) (Bourriez et al., 2023).
• Reliance on synthetic scenarios in training and validation; future work necessitates field deployment and benchmarking in operational environments.
• Coordination protocols sometimes yield oscillatory or symmetric response artifacts in decentralized multi-agent systems (Wartnaby et al., 2019).
• Extension to multi-agent, non-cooperative, or adversarial settings, incorporating richer hybrid physical/digital sensing, game-theoretic constraints, or distributed optimization.
• The integration of perception and planning remains an open area, especially for complex, unstructured scenes, with ongoing work on uncertainty quantification and feedback adaptation.

On-demand collision avoidance thus synthesizes geometric, optimization-based, learning-driven, and biologically inspired control architectures, offering robust safety across domains ranging from satellite constellations and swarms of aerial drones to autonomous vehicles and mobile robots. This diversity of methods and architectures reflects both the generality and domain specificity of challenges in real-time, reactive safety-critical decision making.