Reproduce “CARE: Enhancing Safety of Visual Navigation through Collision Avoidance via Repulsive Estimation” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Joonkyung Kim, Joonyeol Sim, Woojun Kim, Katia Sycara, Changjoo Nam

The CARE framework enhances the safety of existing visual navigation models by integrating monocular depth estimation and repulsive force calculations to adjust trajectories, thereby preventing collisions in out-of-distribution environments. It reduces collision rates by up to 100% and increases collision-free travel distance by up to 10.7 times across diverse robot platforms, operating without additional sensors or fine-tuning.

Key insights

  • Monocular depth estimation can be effectively leveraged to generate local 3D obstacle maps, enabling reactive collision avoidance without dedicated range sensors.
  • An Artificial Potential Field approach, when applied to predicted trajectories, offers a plug-and-play method to steer robots away from obstacles while largely preserving the original navigation policy's intent.
  • A conditional motion control strategy that prioritizes in-place rotation before proceeding (Safe-FOV) helps mitigate risks associated with camera blind spots during turns, ensuring the robot faces a clear path.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.