humanoid loco-manipulation planningExpert
Action-Aligned Latent Planning for Versatile Humanoid Loco-Manipulation
A single planner that steers a frozen whole-body controller through its latent commands reached 42.0% mean success on box pickup, door opening and object grasping in simulation, against 11.6% for replaying the same latents and 3.5% for a kinematic planner, according to a preprint from Amazon FAR, Carnegie Mellon and Stanford. On a real Unitree G1 it succeeded in 6 of 10 box pickups, 4 of 10 door openings and 2 of 10 grasps, trained only in simulation.