Event

Hourglass proposes cross-embodiment motion planning architecture

1 reports · 1 sources

Zhidong Wang and other authors propose the Hourglass architecture, which decomposes cross-embodiment motion planning into two parts: a world model independent of the robot embodiment, and differentiable ego-motion templates for each robot, with the two coupled only through a camera ego-motion interface. The authors state that, in flat-ground scenarios, the robot embodiment affects the planner's observations only through camera ego-motion. On seven simulated robots across five motion types, this general policy's success rate is 3.8–10.6 percentage points higher than imitation learning and diffusion policy baselines on the same world model, matches each robot's specialized policy without dedicated training, and, without using an online map, is 1.4–2.8 percentage points behind a parameter-conditioned MPPI teacher. The authors also state that a model trained on synthetic data, after identification via 3 minutes of teleoperated driving, can zero-shot transfer to a real skid-steer robot unseen in training and navigate in unseen indoor scenes.