[Project website] [Paper] [BFM-Zero upstream]
This repository provides the G1 H200 training source, the selected 166.560M-step checkpoint, and interactive configurations for Carry Box, Push Box, and Kick Box. The G1 task scripts use the same checkpoint. RP1 training configurations and a simulation video are also included.
Click a preview to watch the full video on the project website.
| Carry Box | Push Box | Kick Box |
|---|---|---|
![]() |
![]() |
![]() |
See RP1 details for model and scene settings.
| Path | Contents |
|---|---|
ibfm/ |
Training source and task runtime dependencies |
ibfm/tasks/carry_box/ |
Carry Box interactive script, controller, and latents |
ibfm/tasks/push_box/ |
Push Box interactive script, controller, and latents |
ibfm/tasks/kick_box/ |
Kick Box interactive scripts, controller, and latents |
checkpoints/ibfm-166m/ |
Selected checkpoint and metadata |
deploy/g1_carry_box_166m/ |
G1 Carry Box real-robot deployment, 166.560M ONNX actor, solve-close latents, calibration, and gamepad control |
ibfm/docs/RP1_TRAINING.md |
RP1 official-PD configuration, motion-cache manifest, training commands, and CPU replay |
Linux, Python 3.10, uv, Git LFS, and MuJoCo are required. Install the repository and restore the checkpoint:
git clone https://github.com/Roboparty/I-BFM-Code.git
cd I-BFM-Code
git lfs install
git lfs pull
bash restore_checkpoint.sh
cd ibfm
uv syncAfter setup, run training on eight H200 GPUs from ibfm/:
./scripts/train_ibfm_h200.sh ibfm_reproductionThe launcher uses the published data mix, seed, and training parameters. runs/ibfm_reproduction/checkpoint/ is the rolling resumable checkpoint. policies/latest points to the latest completed policy; policies/best points to the policy with the lowest mean EMD across all 2,719 evaluation motions. These policy copies contain weights and configuration, without optimizer or replay data. The replay buffer is written once, in the final rolling checkpoint, when training finishes normally.
To resume after an interruption:
RESUME=1 ./scripts/train_ibfm_h200.sh ibfm_reproductionWithout a replay buffer from the same step, the continued run will differ from an uninterrupted run.
The training CLI supports --buffer-device cpu to store online replay and expert
buffers in host RAM while keeping the model and training batches on GPU. This
reduces VRAM use for GPUs such as the 4090 48 GB, with additional host memory and
transfer costs. The RP1 launcher accepts BUFFER_DEVICE=cpu, as shown below;
the CLI default remains auto. See memory details.
RP1 uses official PD gains, full domain randomization and LAFAN/Carry/Kick
sampling weights 30/40/30. After preparing the six RP1 caches, run from ibfm/:
PYTHONPATH=. .venv/bin/python scripts/check_rp1_training.py
CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 BUFFER_DEVICE=cpu \
bash scripts/train_rp1.sh rp1_official_pd_250mDefaults are eight GPUs, 512 environments per GPU and 250M total steps. See RP1 training details for data installation, parameters, checkpoint/replay saving, disk requirements and resume commands.
Run these commands from ibfm/ after setup. The scripts create their own output directories for videos, traces, and summaries.
./tasks/carry_box/run_interactive.sh./tasks/push_box/run_interactive.sh./tasks/kick_box/run_interactive.shThe 166.560M-step replay buffer is available on Hugging Face (eight rank-local shards, about 210 GiB total). Download it from ibfm/:
uvx --from huggingface_hub hf download HanSj77/I-BFM-166560M-replay-buffer --repo-type dataset \
--local-dir runs/ibfm_166560m_replayEach runs/ibfm_166560m_replay/checkpoint/buffers/train_rank_0/ through train_rank_7/ directory contains buffer.hdf5 and config.json. The existing TrajectoryDictBufferMultiDim.load() reader in humanoidverse/agents/buffers/trajectory.py loads one shard:
./.venv/bin/python - <<'PY'
from humanoidverse.agents.buffers.trajectory import TrajectoryDictBufferMultiDim
path = "runs/ibfm_166560m_replay/checkpoint/buffers/train_rank_0"
buffer = TrajectoryDictBufferMultiDim.load(path, device="cpu")
print(buffer.size())
PYThis loads the entire selected shard into memory (about 28 GB on disk). The buffer supports latent selection during reward inference; pass a single rank directory to humanoidverse/reward_inference.py with --buffer-path.
To discuss I-BFM, LOGO, reproductions, and humanoid interaction, scan the QR code below to add the author on WeChat.
The training source retains its upstream CC BY-NC 4.0 license. See LICENSE.
If you find this project useful in your research, please cite the I-BFM paper:
@misc{han2026ibfm,
title = {{I-BFM}: Reward-Conditioned Robust Humanoid Interaction via Unsupervised Reinforcement Learning},
author = {Ziqi Han and Yitang Li and Junhan Sun and Fanrong Dong and Yaojie Shen and Lei Ye and Zetong Jing and Yongqi Zhang and Yiming Zhang and Xue Wang and Hao Zhao},
year = {2026},
eprint = {2610.06129},
archivePrefix = {arXiv},
primaryClass = {cs.RO},
url = {https://arxiv.org/abs/2610.06129}
}







