P-BFMPerceptive Behavioral Foundation Modelsfor Humanoid Control via Unsupervised Reinforcement Learning

A single perceptive policy realizes the same intent through distinct behaviors adapted to different terrains.

RoboPartyLab

* Equal contribution† Corresponding author

Same intent.
Different terrain.

P-BFM separates what behavior to perform from how to execute it. Depth perception guides terrain adaptation; explicit heading and velocity instructions guide root motion.

One frozen policy supports motion tracking, goal reaching, reward inference, and teleoperation on a Unitree G1.

How P-BFM works

Interactive simulation

Explore the policy yourself.

Try reference motions, target poses, and reward presets on a shared terrain map. Runs in your browser; desktop recommended.

Open interactive demo

Same motion reference.
Different terrains.

Reference
Upstairs
Slope
Low platforms
Box obstacles
Rough terrain

Teleoperation

Show more videosShow fewer videos

Under perturbations

Push near stairs · Teleoperation
Repeated kicks
Kick near an obstacle
Kick during teleoperation

Goal reaching

Target pose examples

From the interactive demo; not paired with the videos below.

G1 in the low target pose, goal 01
Low pose
G1 lying on its back in target pose 03
Lying down
G1 with both arms raised in target pose 06
Arms raised
Show more videosShow fewer videos

Motion tracking

Show more videosShow fewer videos

Reward inference

The reward specifies forward speed and pelvis-to-foot height.

Velocity target
1 m/s
Height target
0.4 m

Reward targets, not measured outcomes.

Method

From behavioral intent to physical action.

01

Encode the intent

A forward–backward representation maps motion references, target states, and rewards into a shared behavioral space.

02

Perceive the terrain

The actor uses depth observations to adapt whole-body execution to its surroundings.

03

Guide root motion

Explicit Root Instructions specify heading and planar velocity alongside the behavioral latent.

P-BFM pretraining architecture and frozen inference with teleoperation, reference motions, goals, and rewards.
Open full-size figure

Abstract

Behavioral foundation models (BFMs) provide humanoids with reusable repertoires of whole-body behaviors; in real-world humanoid control, the same behavioral intent may require fundamentally different motions under different terrain constraints. Covering such variation with terrain-matched demonstrations requires enumerating a combinatorial space of behaviors and terrain geometries.

We introduce P-BFM, a perceptive behavioral foundation model that factorizes humanoid control into terrain-invariant intent and terrain-conditioned realization. A forward–backward representation maps motion references, target states, and reward functions into a shared intent space without terrain information, while a perceptive policy learns how to realize each intent under the surrounding terrain geometry. We further identify a complementary limitation of purely latent control: behavioral latents can preserve local motion patterns while losing precise global progress. We therefore introduce Explicit Root Instructions (ERI), which separates global transport (e.g., heading and planar velocity) from behavioral intent, while leaving local whole-body adaptation to the perceptive policy.

This decomposition leads to a characteristic capability absent from conventional BFMs: one fixed intent induces different, terrain-appropriate behaviors, including changes in step height, stride, posture, foothold placement, and contacts, without terrain-specific demonstrations. Across challenging terrains, P-BFM improves motion fidelity, root-control precision, and perturbation recovery over strong baselines. A single frozen policy further supports motion tracking, goal reaching, reward inference, teleoperation, and terrain-aware recovery, and transfers zero-shot to a Unitree G1 using onboard depth sensing.