Copper-Policy

Focus on the Representation for Robust Robot Manipulation

Zexin Feng · Yixu Feng · Lingyu Xiao · Shang Su · Kexin Zheng
Chang Xu · Mengkai Shi · Shuo Feng · Xintao Yan

The University of Hong Kong · The University of Sydney · Tsinghua University · DenseAI

English | 中文

Official inference checkpoints for Copper-Policy: Focus on the Representation for Robust Robot Manipulation.

Copper-Policy combines a compact world representation with spatial visual features for robot manipulation. This release contains two approximately 2B-parameter policies.

📄 Paper · 🌐 Project Page · 💻 Code

📦 Included checkpoints

libero/
├── policy.pt
└── policy.json
robotwin/
├── policy.pt
└── policy.json
release_manifest.json
SHA256SUMS
LICENSE

• LIBERO: one checkpoint shared by LIBERO and LIBERO-Plus; 2 RGB views, 7-dimensional actions and 8-dimensional proprioception.
• RoboTwin: one checkpoint shared by clean and randomized evaluations; 3 RGB views, 14-dimensional actions and proprioception.
• Both checkpoints predict an action horizon of 32. Normalization statistics and the configuration needed for inference are included in policy.pt; no separate dataset-statistics file is required.
• policy.json describes each checkpoint; SHA256SUMS provides checksums. These custom PyTorch checkpoints are loaded using the Copper-Policy code repository.

📥 Download

Checkpoint	Benchmarks
libero/policy.pt	LIBERO / LIBERO-Plus
robotwin/policy.pt	RoboTwin clean / randomized

Each checkpoint is approximately 4.06 GB. Both together require approximately 8.12 GB.

# Run in the Copper-Policy code repository
uv run --frozen --all-extras --no-sync hf download Mark455/Copper-Policy \
  libero/policy.pt robotwin/policy.pt \
  --local-dir pretrained_weights/copper_policy

To download only one model, keep only its file path in the command.

For installation and inference, see the code repository.

⚙️ Inference

The code repository provides four evaluation launchers:

uv run --frozen --all-extras --no-sync bash test_libero.sh --all-tasks
uv run --frozen --all-extras --no-sync bash test_libero_plus.sh --all-tasks
uv run --frozen --all-extras --no-sync bash test_robotwin_clean.sh --all-tasks
uv run --frozen --all-extras --no-sync bash test_robotwin_rand.sh --all-tasks

Default inference uses 10 denoising steps, with 10 replan steps for LIBERO / LIBERO-Plus and 24 for RoboTwin. Compilation and rollout videos are enabled. Download the V-JEPA 2.1, DINOv2 and Wan T5 encoder presets using python -m tools.download_weights all --yes through the repository's uv environment. No separate Wan transformer initialization weights are needed.

Follow the code repository for installation and simulator assets; adapt PyTorch/CUDA to your hardware and driver. Training code — coming soon.

Note: LIBERO language mapping

LIBERO training instructions sometimes differ from benchmark prompts. The code maps exact plaintext benchmark prompts to training instructions. LIBERO-Plus language perturbations are not mapped; unmatched prompts retain the benchmark text.

📊 Reported results

LIBERO	LIBERO-Plus	RoboTwin clean	RoboTwin randomized
97.25%	80.85%	70.84%	12.98%

These are results reported in the paper and project page. Individual evaluation runs record their own measured scores.

📚 Citation

@misc{feng2026copperpolicyfocusrepresentationrobust,
  title={Copper-Policy: Focus on the Representation for Robust Robot Manipulation},
  author={Zexin Feng and Yixu Feng and Lingyu Xiao and Shang Su and Kexin Zheng and Chang Xu and Mengkai Shi and Shuo Feng and Xintao Yan},
  year={2026},
  eprint={2609.32779},
  archivePrefix={arXiv},
  primaryClass={cs.RO},
  url={https://arxiv.org/abs/2609.32779}
}

🙏 Acknowledgements and license

Our implementation builds on FastWAM and LingBot-VA. Dataset conversion uses our lerobot-tools. Codex helped organize the open-source code.

Released under Apache 2.0. External encoders and simulator assets retain their respective licenses.