MinBC: Minimal Behavior Cloning

MinBC is a lightweight library for training behavior cloning (BC) policies for robot manipulation. It has been used in several of our research projects and supports multiple imitation learning algorithms with a unified training interface.

Currently supported policies:

• Standard (Vanilla) Behavior Cloning with action chunking
• Diffusion Policy
• Choice Policy (see our paper for details)

Disclaimer: This codebase is under active development. For the reference implementation used in the dexscrew paper, please refer to this version.

Installation

Prerequisites

• Python 3.8+
• CUDA-compatible GPU

Install Dependencies

conda create -n minbc python=3.11
conda activate minbc
pip install -r requirements.txt

Quick Start

This section demonstrates how to train a policy using a small humanoid manipulation dataset from our Choice Policy paper.

Dataset Preparation

Download the gr1_dishwasher_debug dataset

gdown 1iQJRAC0CMp4P2UhDFx9o1tBmG4bGfGkF -O data/gr1_dishwasher_debug.zip
cd data
unzip gr1_dishwasher_debug.zip -d ./
rm gr1_dishwasher_debug.zip
cd ../

It contains 1 training trajectory and 1 test trajectory. The expected directory structure is:

data/
  gr1_dishwasher_debug/
    train/
    test/

Training

Using this minimal data, you can run the following command for training and evaluation:

python train.py train --gpu 0 --seed 0 \
--optim.num_epoch 100 --optim.learning-rate 0.0005 --optim.batch-size 64 \
--data.pred-head-act \
--policy-type bc --dp.action_decoder cond_hourglass --data.base_action_dim 26 \
--data.data-key img gr1_upper hand_qpos --data.im-key head_image left_wrist_image right_wrist_image \
--data.im-encoder scratch \
--train_data gr1_dishwasher_debug/train --test_data gr1_dishwasher_debug/test \
--output_name gr1_dishwasher_debug/example_test/

In the above command, num_epoch, learning_rate, batch_size are standard optimization parameters. train_data, test_data, output_name are training path setup. policy-type specifies which imitation learning policy you want to use, use policy-type bc for naive behavior cloning, policy-type dp for diffusion policy, --policy-type bc --dp.num-proposl 5 for choice policy.

Another useful setup parameter is data-key and im-key, as the name indicates, it specify what input modality is used. If your dataset has another input type, you should also manually add the vector dimension for the INPUT_DIM variable in config/base.py.

For multi-gpu training, you can use the following command OMP_NUM_THREADS=1 torchrun --standalone --nnodes=1 --nproc_per_node=4 learning/train.py train --gpu 0,1,2,3 --multi-gpu For using pretrained image encoder one can use --data.im-encoder DINOv3 --dp.encoder.im-encoder-frozen

OMP_NUM_THREADS=1 torchrun --standalone --nnodes=1 --nproc_per_node=4 learning/train.py train --gpu 0,1,2,3 --multi-gpu --seed 3 \
--optim.num_epoch 100 --optim.learning-rate 0.0005 \
--data.pred-head-act --optim.batch-size 128 \
--policy-type bc --dp.action_decoder cond_hourglass \
--train_data 250815_dishwasher_nowaist/train \
--test_data 250815_dishwasher_nowaist/test \
--data.im-encoder DINOv3 --dp.encoder.im-encoder-frozen --data.data-key img gr1_upper hand_qpos \
--output_name 250815_dishwasher_nowaist/250827_all_bc_without_clip_3

Data Format

data/
└── your_dataset/
    ├── train/
    │   ├── episode_000/
    │   │   ├── step_000.pkl
    │   │   └── ...
    │   └── ...
    └── test/
        └── ...

Each .pkl file should contain:

• action: numpy array (action_dim,) - required
• joint_positions, joint_velocities, xhand_pos, etc.: numpy arrays - optional
• base_rgb: numpy array (H, W, 3) - optional, for vision

Training Output

Results saved to outputs/<output_name>/:

• config.json - Training configuration
• model_best.ckpt - Best model checkpoint
• stats.pkl, norm.pkl - Normalization parameters

Monitor with: tensorboard --logdir outputs/

Citations

@article{hsieh2025learning,
  title={Learning Dexterous Manipulation Skills from Imperfect Simulations},
  author={Hsieh, Elvis and Hsieh, Wen-Han and Wang, Yen-Jen and Lin, Toru and Malik, Jitendra and Sreenath, Koushil and Qi, Haozhi},
  journal={arXiv:2512.02011},
  year={2025}
}

@article{qi2025coordinated,
  title={Coordinated Humanoid Manipulation with Choice Policies},
  author={Qi, Haozhi and Wang, Yen-Jen and Lin, Toru and Yi, Brent and Ma, Yi and Sreenath, Koushil and Malik, Jitendra},
  journal={arXiv:2512.25072},
  year={2025}
}

Questions?

Contact Yen-Jen Wang or Haozhi Qi.

Acknowledgements

MinBC is modified from HATO DP part, which is a simplification of the original Diffusion Policy.