I-BFM: Reward-Conditioned Robust Humanoid Interaction via Unsupervised Reinforcement Learning

Ziqi Han*

Yitang Li*

Junhan Sun*

Fanrong Dong

Yaojie Shen

Lei Ye

Zetong Jing

Yongqi Zhang

Yiming Zhang

Xue Wang

Hao Zhao

PDF Cite Code Source Document DOI

I-BFM project preview showing reward-conditioned humanoid-object interaction and a carrying behavior.

I-BFM project preview, from the authors’ project website (RoboParty Lab).

Abstract

I-BFM extends behavioral foundation models to humanoid-object interaction. Unsupervised reinforcement learning and forward-backward representations learn a shared behavior space for the robot, objects, and their contacts. A downstream reward is mapped to a latent command for the same policy, avoiding task-specific policy optimization. Training combines short-horizon interaction targets and longer-horizon goals. The policy supports carrying, pushing, kicking, and task chaining, with simulation studies and Unitree G1 hardware demonstrations of interaction and recovery from failures and disturbances.

Type

Manuscript

Publication

arXiv preprint arXiv:2610.06129 (2026)

Equal contribution: Ziqi Han, Yitang Li and Junhan Sun.

One behavior space for interaction

I-BFM learns a shared representation of humanoid motion, object motion, and contact through unsupervised reinforcement learning. Downstream rewards are converted into latent commands, allowing one policy to perform different interaction tasks without optimizing a new task-specific policy.

Robust execution beyond nominal motion

Short-horizon interaction targets and longer-horizon goals support carrying, pushing, kicking, and task chaining. The work evaluates recovery after falls and external disturbances, alongside hardware loco-manipulation on a Unitree G1. The reported simulation Carry success is 94.3% under nominal conditions and 89.3% after robot falls in the evaluated benchmark.

I-BFM training and reward-inference overview

The overview separates representation pretraining from downstream reward inference. Project videos provide the corresponding simulation and real-robot demonstrations.

Project and demonstrations · Preprint · Code

Last updated on Oct 13, 2026