NYU's RCSP algorithm prevents robot collisions by predicting near-misses

Mobile robots often fail not because of an immediate collision, but because a currently safe velocity commits them to a path that moving obstacles will soon block. Han and Zhu's RCSP addresses this predictive near-miss commitment problem by evaluating candidate commands against plausible short-horizon obstacle futures. The algorithm maintains a lightweight probabilistic belief over motion conjectures, simulates future interactions, penalizes high-risk trajectories, and executes a local safety check before acting. In controlled MuJoCo bottleneck scenarios, RCSP reached the goal without collisions and showed higher secondary safety and path quality compared to a non-adaptive predictor, albeit with additional latency.

When integrated into a ROS2/Gazebo environment as a safety layer atop the standard Nav2 stack, RCSP significantly reduced dynamic near-miss failures. However, on the official DynaBARN/Jackal transfer benchmark, well-tuned DWA and TEB planners still outperformed RCSP on strict success metrics, highlighting the current boundaries of the approach. The authors position RCSP as a complementary predictive-risk module for dynamic bottleneck regimes rather than a full replacement. This work offers a practical way to enhance existing navigation stacks with risk-aware foresight.