Humanoid loco-manipulation demands a tight coupling between high-level task choices and low-level whole-body control. The researchers argue that the representation used to bridge these levels is crucial: kinematic or contact targets are compact but miss interaction-specific control responses, while direct low-level action prediction burdens the planner with high-frequency feedback.
To address this, they introduce FIBER, which adopts the executable latent space of a frozen behavioral foundation model (BFM) as a shared planning representation. A latent policy projection step aligns independently trained task policies with this space by finding latent commands whose induced actions match local control responses. This enables lower-rate planning while preserving the BFM's high-frequency state feedback. The team also presents AH-BFM, an adaptive-horizon BFM trained with multi-gamma forward-backward learning to handle both precise control and extended whole-body behavior.
In simulation, FIBER was evaluated on box pickup, door opening, and object grasping across 128 variations per task, reaching a mean success of 42.0%. This more than tripled the projected latent replay baseline (11.6%) and substantially outperformed kinematic planning (3.5%). On a real Unitree G1 humanoid, the same unified planner demonstrated sim-to-real transfer, skill chaining, and conditional steering.