Mitigating Multi-Module Errors for Reliable Navigation in Dynamic Environments via Online Trajectory Refinement

QI YIN, Yangming Shi

Keywords: Quadruped Robot Navigation

Abstract: Perception in dynamic obstacle avoidance scenarios is inevitably affected by errors from multiple perception modules, e.g., robot localization, dynamic detection, and trajectory prediction. Existing methods mainly focus on reducing errors in a single module, overlooking their interactions and accumulation, which degrade navigation performance. We propose a systematic framework to mitigate multi-perception module errors and improve the reliability of dynamic obstacle avoidance. The proposed framework integrates two key innovations: an online trajectory refinement network called Residual-DtACI, and a point cloud–based dynamic target localization method. The Residual-DtACI adaptively refines trajectories through multi-expert residual estimation, while the point cloud–based method improves localization accuracy and robustness to occlusions and pose variations. We conducted several experiments to evaluate the proposed framework regarding its effectiveness in reducing accumulated perception errors. Results found that errors across different perception modules were reduced, alleviating error accumulation and enhancing prediction accuracy. In addition, the evaluations on public benchmarks show significant gains, with Residual-DtACI improving robot trajectory accuracy by up to 88.5%, and the point cloud–based method increasing localization accuracy by 59.9%.

Email Sharing: We authorize the sharing of all author emails with Program Chairs.

Data Release: We authorize the release of our submission and author names to the public in the event of acceptance.

Submission Number: 8