Humanoid robots operating in human-centric environments must coordinate the entire body to traverse tight, cluttered spaces while maintaining balance. According to the paper's abstract, existing methods often struggle to jointly consider dynamic stability and collision avoidance, leaving a gap in practical whole-body navigation.
The authors propose SWAP, a test-time steerable method for whole-body collision avoidance. Rather than treating the robot as a simple shape, SWAP is designed to coordinate the full body during motion, which is essential for squeezing through narrow passages or avoiding obstacles in close quarters.
Because it operates at test time, SWAP could potentially be applied to existing control policies without retraining, though the abstract does not provide implementation details. The work is described in arXiv:2610.25711.