FIBER: Action-Aligned Latent Planning for Humanoid Loco-Manipulation

Loading video

Loading video

FIBER turns a frozen Behavior Foundation Model's executable latent space into the shared planning interface: Latent Policy Projection maps independently trained skill experts into latent commands whose induced actions match their local responses, then a single perceptive flow planner is trained entirely offline on their closed-loop latent rollouts. Across 128 simulated variations per task it reaches 42.0% mean success, against 11.6% for projected latent replay and 3.5% for kinematic planning, and transfers to a real Unitree G1 for box pickup, door opening and object grasping.