Manuscript | Raymon Yip CV

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

Mon, 05 Oct 2026 00:00:00 +0000

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.

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

· ·

PredActor: Predictive Action Diffusion for Steerable Onboard Humanoid Control

Mon, 21 Sep 2026 00:00:00 +0000

From motion generation to direct control

PredActor addresses the gap between flexible motion generation and closed-loop execution. A separate reference tracker is replaced by a policy that generates actions directly and uses its own future-state predictions as a guidance interface. Proprioceptive feedback closes the loop as inference rolls forward.

Steering one policy

Text-conditioned behavior: classifier-free guidance strengthens the learned motion condition.
Test-time objectives: classifier guidance steers internal predicted states, enabling joystick control and other motion objectives.
Hardware execution: rolling denoising and computation-preserving runtime optimizations support a 20 ms control period on the onboard Jetson Orin NX.

The paper reports a measured complete callback latency of 16.790 ms median and 19.383 ms p95 on the Jetson Orin NX. Only generated actions are applied to the robot; predicted states remain inside the policy for guidance.

· ·