Reactive Composition for Multi-Goal Robotic Tasks

This research presents a reactive control approach for multi-goal robotic tasks that dynamically resolves conflicts in goal prioritization during execution. By extending the AICON framework with adaptive nullspace projections, the method allows gradients to interact based on current magnitudes rather than fixed hierarchies. It demonstrates robust performance on 100 non-convex navigation and 100 pushT problems, outperforming static potential fields and diffusion policies. The approach also handles perceptual uncertainty, joint limits, and self-collisions on real robots, achieving success in 49 of 50 pushing trials.