Event

FIBER framework enables humanoid robots to complete multiple manipulation tasks with a unified planner

1 reports · 1 sources

Yutang Lin, Guanya Shi, and C. Karen Liu propose the FIBER framework, which uses the executable latent space of a frozen behavior foundation model (BFM) as a shared planning representation for heterogeneous interaction skills, maps independently trained task policies into this representation via latent policy projection, achieves low-frequency planning while retaining the BFM's high-frequency state feedback, and then trains a single perception-stream planner offline; they also propose an adaptive-horizon BFM (AH-BFM) trained with multi-gamma forward-backward learning. In simulation, on three task types—box picking, door opening, and object grasping—each tested across 128 variations, FIBER achieves an average success rate of 42.0%, more than three times that of projected latent space replay (11.6%) and markedly higher than kinematic planning (3.5%); experiments on a real Unitree G1 demonstrate sim-to-real transfer, skill chaining, and conditional turning, all completed by the same unified planner.