The research system samples full-body maneuvers at deployment, then favors those matching a chosen goal.
The authors posted SWAP, or Steerable Whole-body Avoidance Policy, as an arXiv research paper on October 19. They tested it in simulation and on a physical Unitree G1; it is not a product available to buy or a commercial deployment.
SWAP divides control into two layers. A high-level diffusion policy—a model that generates possible actions—proposes whole-body motions from the robot’s sensor data. Those options can include ducking, sidestepping, or twisting. A lower-level reinforcement-learning controller tracks the selected motion while maintaining balance.
At deployment, SWAP uses Monte Carlo sampling: it creates multiple candidate motion chunks, scores them, and selects the best one. The score can favor speed, heading, body height, or posture. That lets a user or developer change the robot’s preference without retraining the main policy for each objective.
The authors train collision avoidance through “hindsight labeling.” In simulation, they execute randomized motions, calculate the space swept by the robot’s body, place obstacles outside that space, and recreate what the robot would have sensed. This automatically produces collision-free training examples.
The paper reports an 88.3% average success rate across five simulated obstacle tasks. In physical tests, the G1 succeeded four of five times walking sideways, three of five ducking, four of five lifting a hand, three of five against a fast-moving obstacle, and four of five against a slow one.
The authors also report a 34% average improvement in targeted rewards such as speed, orientation, and posture. During real-world tests, LiDAR supplied obstacle perception; motion capture localized the robot and specified dynamic-obstacle states. It did not provide static-obstacle perception.
Failures mainly involved LiDAR blockage during extreme poses and differences between simulated and real motion tracking. More varied, cluttered scenes remain the next practical test.