P-BFM: Perceptive Behavioral Foundation Models for Humanoid Control via Unsupervised Reinforcement Learning

English | 中文

    

P-BFM is a perceptive behavioral foundation model for terrain-adaptive humanoid control. It maps motion references, target body states, and reward functions into a shared intent space. A single frozen policy combines these intents with depth perception and explicit heading and velocity instructions to execute whole-body behaviors across terrains.

Training uses 40 flat-ground LaFAN motion sequences and online interaction with simulated terrains, without terrain-specific motion demonstrations. The policy transfers zero-shot to a Unitree G1 with onboard depth sensing, supporting motion tracking, goal reaching, reward inference, teleoperation, and terrain-aware recovery.

Architecture · Training · Inference / export

Layout

P-BFM/
├── pbfm/                  Runtime package
│   ├── train.py           Training CLI
│   ├── configs/           One home for model, simulator, robot and depth settings
│   ├── agents/            Actor, FB, discriminator, Q critics and replay
│   ├── training/          Collection, updates, checkpointing and training evaluation
│   ├── envs/              MJLab bridge, observations, resets and physical rewards
│   ├── perception/        Depth processing and camera utilities
│   ├── terrains/          Terrain generation and curriculum
│   ├── motion/            Motion loading, conversion and reference observations
│   ├── robots/            G1 asset parsing and validation
│   ├── inference/         Tracking / goal / reward; shared runtime and ONNX export
│   ├── evaluation/        Offline metrics, terrain comparisons and diagnostics
│   ├── data/              Bundled G1 assets and a small CSV example
│   └── utils/             Shared numerical and resource helpers
├── scripts/               Launch, download, export and verification commands
├── tests/                 Regression tests and their small fixtures
└── docs/                  Usage and architecture

Configuration files are bundled under pbfm/configs/, including in the installed package; downloaded motions, weights, replay and videos stay external.

Model and settings

Component	Hidden width × layers
Actor / Forward F	2048 × 6
QD	1024 × 4
QAux	2048 × 4
QH / QV	1024 × 3

The four critics are independent; each uses two Q estimates. The training preset is fb_depth: BCE/log-odds, D=.01, Aux=.02, QH=QV=.002, dynamic Aux normalization, LaFAN-only, slippage=-4, stumble=-2. Depth is Crop18 (8×18×32), CNN stride 1, frame feature 256 and GRU feature 256. The velocity estimator is depth-supervised.

Install and train

Python 3.10, MJLab 1.4, MuJoCo 3.8, PyTorch 2.7. Install from the repository root:

uv sync --extra test
# Optional: download the separately hosted LaFAN files.
bash scripts/download_data.sh

Fresh 192M training (eight free GPUs, GPU replay; edit the data/output paths):

export PYTHONPATH="$PWD${PYTHONPATH:+:$PYTHONPATH}"
export OMP_NUM_THREADS=1 MUJOCO_GL=egl NCCL_NVLS_ENABLE=0
export UFO_CACHE_DIR=/path/to/writable-disk/pbfm-cache

uv run python -m torch.distributed.run --standalone --nnodes=1 --nproc_per_node=8 \
  scripts/train_legacy_scratch.py --gpu-ids single --terrain-mode rp1_simple \
  --num-envs 1024 --num-env-steps 192020480 --seed 4831 \
  --data-path /path/to/lafan_29dof_10s-clipped.pkl \
  --work-dir /path/to/new-run

Add --use-wandb after authenticating separately. Add --smoke with a different new work directory for a 32768-step short test. These commands do not stop other users' jobs or wait for free GPUs.

Use the scratch entry point above for training from step 0: it sets 192020480 total global steps and starts the terrain curriculum at step 0. See Training for schedule and checkpoint requirements.

Inference and export

uv run python -m pbfm.inference.tracking --help
uv run python -m pbfm.inference.goal --help
uv run python -m pbfm.inference.reward --help
uv run python -m pbfm.evaluation.terrain --help
uv run python -m scripts.export_depth_policy_onnx --help
uv run python -m scripts.export_backward_encoder_onnx --help

Tracking selects terrain and camera settings in the same entry point. The policy ONNX includes depth encoding and velocity estimation; the backward encoder is exported separately. Input normalization, context and hardware control requirements are described in the inference guide.

Tests

PYTHONPATH=. uv run python -m pytest tests

External-data/hardware checks are explicitly skipped when their inputs are unavailable. CPU and small-model ONNX tests are not full training or real-robot validation.

PBFM builds on UFO. Original notices and the upstream CC BY-NC 4.0 license are retained.

Project and community

• Project website and demos: https://roboparty.github.io/P-BFM/
• Paper: https://arxiv.org/abs/2610.15833
• Team: RoboPartyLab
• Contact: xuewangusst@gmail.com
• To join our discussion group, add 15251239451 on WeChat.

Repository branches

• main: PBFM training, evaluation and inference for Unitree G1.
• pages: project website, paper, figures, and demonstration videos. GitHub Pages publishes from the root of this branch.

Website updates should be committed to pages, independently of code changes on main.