---
license: mit
tags:
- robotics
- imitation-learning
- robot-manipulation
- long-horizon-manipulation
- non-markovian-control
- rmbench
- chronos
- pytorch
---

# Chronos RMBench Checkpoints

This repository provides official Chronos checkpoints and matching normalization files for RMBench experiments.

Chronos is a physics-informed full-history imitation learning framework for non-Markovian long-horizon robot manipulation.

Links:

- [Project Page](https://chronos-manipulation.github.io/)
- [arXiv Paper](https://arxiv.org/abs/2606.30318)
- [GitHub Code](https://github.com/yulinzhouZYL/Chronos)

---

## Files

This repository contains checkpoint and scaler files for multiple RMBench tasks.

For each task, the checkpoint and scaler must be used together:

```text
checkpoint file: last.ckpt
scaler file    : scaler_<task_name>_ee_3d.pth
```

For example, for `cover_blocks`:

```text
last.ckpt
scaler_cover_blocks_ee_3d.pth
```

If a checkpoint is downloaded with a task-specific filename, please rename it to `last.ckpt` after placing it in the expected local checkpoint directory.

---

## Action Contract

These checkpoints use the Chronos RMBench EE-pose action representation.

The action is a 16D dual-arm EE pose + gripper vector:

```text
left EE pose  : x, y, z, qw, qx, qy, qz
left gripper  : 1 dimension
right EE pose : x, y, z, qw, qx, qy, qz
right gripper : 1 dimension
```

The policy should be evaluated with RMBench EE-pose control:

```python
TASK_ENV.take_action(action, action_type="ee")
```

These are not qpos replay checkpoints.

---

## Expected Local Layout

Please place each downloaded checkpoint and scaler in the Chronos repository as follows:

```text
RMBench/policy/Chronos/
├── scaler_<task_name>_ee_3d.pth
└── checkpoints/
    └── <task_name>/
        └── EE_16/
            └── last.ckpt
```

For example, for `cover_blocks`:

```text
RMBench/policy/Chronos/
├── scaler_cover_blocks_ee_3d.pth
└── checkpoints/
    └── cover_blocks/
        └── EE_16/
            └── last.ckpt
```

---

## Evaluation

Before evaluation, make sure `RMBench/policy/Chronos/deploy_policy.yml` points to the correct checkpoint and scaler.

For `cover_blocks`, use:

```yaml
ckpt_path: "policy/Chronos/checkpoints/cover_blocks/EE_16/last.ckpt"
scaler_path: "policy/Chronos/scaler_cover_blocks_ee_3d.pth"
temporal_agg: true
gpu_id: 0
```

Then run:

```bash
cd RMBench/policy/Chronos
bash eval.sh cover_blocks demo_clean Chronos 42 0
```

For another task, replace `cover_blocks` with the corresponding RMBench task name and use the matching scaler:

```yaml
ckpt_path: "policy/Chronos/checkpoints/<task_name>/EE_16/last.ckpt"
scaler_path: "policy/Chronos/scaler_<task_name>_ee_3d.pth"
temporal_agg: true
gpu_id: 0
```

---

## Important Notes

- The checkpoint and scaler must come from the same task.
- The scaler filename should follow the pattern `scaler_<task_name>_ee_3d.pth`.
- These checkpoints are for EE-pose control, not qpos replay.
- Please use the Chronos-compatible `demo_clean.yml` included in the GitHub repository.
- If offline validation looks reasonable but closed-loop evaluation fails, first check whether `deploy_policy.yml` points to the correct scaler.

---

## Citation

If you find Chronos useful, please cite:

```bibtex
@article{zhou2026chronos,
  title={Chronos: A Physics-Informed Full-History Framework for Non-Markovian Long-Horizon Manipulation},
  author={Zhou, Yulin and Wang, Yimeng and Wang, Nengyu and Xing, Shaojia and Tu, Shiyun and Li, Xiang and Zhang, Jingkai and Jiang, Ningbo and Lin, Yuankai and Yang, Hua and Zeng, Xiangrui and Yin, Zhouping},
  journal={arXiv preprint arXiv:2606.30318},
  year={2026}
}
```

---

## License

This checkpoint repository is released for research use.

RMBench benchmark components follow the original RMBench license.