🔋 EnergyAction: Unimanual to Bimanual Composition with Energy-Based Models (CVPR 2026)

Mingchen Song^{1 2}, Xiang Deng^{1 3†}, Jie Wei¹, Dongmei Jiang², Liqiang Nie¹, Weili Guan^{1 3†}

¹Harbin Institute of Technology, Shenzhen   ²PengCheng Laboratory   ³Shenzhen Loop Area Institute
^†Corresponding author

If you find this work useful for your research, please kindly cite our paper and star our repo.

📢 Updates

• [03/2026] 🔥 Code and pretrained models are released.
• [03/2026] 📄 arXiv paper released.](https://arxiv.org/abs/2603.20236)

Introduction

We propose EnergyAction, a framework that compositionally transfers pre-trained unimanual manipulation policies to bimanual tasks via Energy-Based Models (EBMs). By modeling each arm's policy as an energy function and composing them through energy summation, EnergyAction generates coordinated dual-arm actions with temporal coherence and collision avoidance — requiring only minimal bimanual demonstration data.

Installation

1. Clone this repository

git clone https://github.com/xxx/EnergyAction.git
cd EnergyAction

2. Create conda environment

conda create -n energyaction python=3.10 -y
conda activate energyaction
pip install -e .

3. Install additional packages

pip install torch torchvision open_clip_torch zarr einops setproctitle

4. Install CoppeliaSim (required for evaluation)

Online evaluation requires CoppeliaSim and PyRep. Follow the PyRep installation guide to set up CoppeliaSim, then set the environment variables:

export COPPELIASIM_ROOT=/path/to/CoppeliaSim
export LD_LIBRARY_PATH=${LD_LIBRARY_PATH}:${COPPELIASIM_ROOT}
export QT_QPA_PLATFORM_PLUGIN_PATH=${COPPELIASIM_ROOT}

Datasets

We use the RLBench benchmark with Peract2 data format stored in Zarr for efficient loading.

Single-arm data (RLBench)

Used for Stage 1: pretraining the single-arm 3D Flow Actor backbone.

# Download single-arm RLBench training data
wget https://huggingface.co/katefgroup/3d_diffuser_actor/resolve/main/Peract_packaged.zip
unzip Peract_packaged.zip
rm Peract_packaged.zip

# Download single-arm RLBench evaluation data
wget https://huggingface.co/katefgroup/3d_flowmatch_actor/resolve/main/peract_test.zip
unzip peract_test.zip
rm peract_test.zip

# Convert to Zarr format
python data_processing/peract_to_zarr.py \
    --root <raw_single_arm_data_dir> \
    --tgt <zarr_output_dir>

Bimanual data (RLBench2)

Used for Stage 2: training the EBM Bimanual Composer.

# Download bimanual-arm RLBench data (Peract format)
python scripts/rlbench/download_peract2.py --root <data_dir>

# (Optional) Generate data from scratch
bash data_generation/generate_data.sh

# Convert raw bimanual data to Zarr format
python data_processing/peract2_to_zarr.py \
    --root <raw_bimanual_data_dir> \
    --tgt <zarr_bimanual_output_dir>

Pretrained Checkpoints

We release all pretrained checkpoints on Hugging Face: 🤗 mingchens/EnergyAction

# Download checkpoints (requires huggingface_hub)
pip install huggingface_hub

# Download all checkpoints
huggingface-cli download mingchens/EnergyAction --local-dir checkpoints/

# Or download individually
huggingface-cli download mingchens/EnergyAction pretrain_unimanual_checkpoint.pth --local-dir checkpoints/
huggingface-cli download mingchens/EnergyAction 100demo_checkpoint.pth --local-dir checkpoints/

• pretrain_unimanual_checkpoint.pth: Use as LEFT_PRETRAINED_PATH and RIGHT_PRETRAINED_PATH in Stage 2 training. You can skip Stage 1 entirely with this checkpoint.
• 20demo_checkpoint.pth: EBM Bimanual Composer trained with 20 bimanual demonstrations per task.
• 100demo_checkpoint.pth: EBM Bimanual Composer trained with 100 bimanual demonstrations per task.

Evaluation

Online Evaluation (RLBench + CoppeliaSim)

We evaluate on 13 bimanual tasks in RLBench using CoppeliaSim simulation:

# Evaluate with experiment name (auto-finds best.pth)
bash scripts/rlbench/eval_ebm_bimanual_composer.sh my_experiment

# Evaluate with specific checkpoint
bash scripts/rlbench/eval_ebm_bimanual_composer.sh /path/to/checkpoint.pth

# With nohup (recommended for long evaluation)
nohup bash scripts/rlbench/eval_ebm_bimanual_composer.sh my_experiment > eval_my_experiment.log 2>&1 &

Project Structure

EnergyAction/
├── main.py                            # Entry point
├── setup.py                           # Package installation
├── modeling/
│   ├── ebm_compositionality/          # 🔑 Core EBM modules
│   │   ├── flow_to_energy.py          #   Flow Matching → Energy conversion
│   │   ├── energy_composer.py         #   E_left + E_right + E_coord composition
│   │   ├── ebm_bimanual_composer.py   #   Main EBM bimanual model
│   │   ├── bimanual_coordination_constraints.py  # Physics-informed constraints
│   │   └── panda_kinematics.py        #   Franka Panda robot kinematics
│   ├── encoder/                       # Vision-language encoders (CLIP backbone)
│   ├── noise_scheduler/               # DDPM / DDIM / Rectified Flow
│   ├── policy/                        # Base denoising policy networks
│   └── utils/                         # Transformer layers, attention, PE
├── datasets/                          # RLBench data loading (Zarr format)
├── utils/                             # Training loop, schedulers, EMA
├── pretrain_unimanual/                # Pretrained single-arm checkpoint
├── scripts/
│   └── rlbench/
│       ├── train_single_arm.sh        # Stage 1: Single-arm training
│       ├── train_ebm_bimanual_composer.sh  # Stage 2: EBM bimanual training
│       └── eval_ebm_bimanual_composer.sh   # Online evaluation
├── online_evaluation_rlbench/         # CoppeliaSim online evaluation
├── data_processing/                   # Raw data → Zarr conversion
├── data_generation/                   # RLBench data generation
└── instructions/                      # Task language instructions

Acknowledgement

Our implementation is mainly based on the following codebases. We gratefully thank the authors for their wonderful works.

• 3D Diffuser Actor (3DFA)
• Peract2

Citation

@article{energyaction2026,
  title={EnergyAction: Energy-Based Compositional Flow Matching for Bimanual Manipulation},
  author={Song, Mingchen and Deng, Xiang and Wei, Jie and Jiang, Dongmei and Nie, Liqiang and Guan, Weili},
  journal={arXiv preprint arXiv:2603.20236},
  year={2026}
}