README.md

---
license: apache-2.0
pipeline_tag: robotics
tags:
- robotics
- manipulation
- rlbench
- imitation-learning
---

# SkiP — Pretrained Checkpoints (RLBench-10)

Pretrained checkpoints for **SkiP (Skip Policy)** on 10 RLBench tasks, as presented in the paper [SkiP: When to Skip and When to Refine for Efficient Robot Manipulation](https://arxiv.org/abs/2605.15536).

SkiP dynamically leaps over redundant free-space steps and concentrates control on contact-rich key segments via action relabeling — no learned skip planner or hierarchical structure required.

## Links

- **Paper:** [https://arxiv.org/abs/2605.15536](https://arxiv.org/abs/2605.15536)
- **Code:** [https://github.com/CCCalcifer/Skip-Policy](https://github.com/CCCalcifer/Skip-Policy)
- **Project Page:** [https://pgq18.github.io/SkiP-page/](https://pgq18.github.io/SkiP-page/)

## Checkpoints

| File | Task |
|------|------|
| checkpoints/open_box.pt | open_box |
| checkpoints/open_drawer.pt | open_drawer |
| checkpoints/pick_up_cup.pt | pick_up_cup |
| checkpoints/press_switch.pt | press_switch |
| checkpoints/push_button.pt | push_button |
| checkpoints/reach_target.pt | reach_target |
| checkpoints/stack_wine.pt | stack_wine |
| checkpoints/sweep_to_dustpan.pt | sweep_to_dustpan |
| checkpoints/take_lid_off_saucepan.pt | take_lid_off_saucepan |
| checkpoints/turn_tap.pt | turn_tap |

**Training config:** kf_mode=skip, skip.window=16, skip.high_quantile=0.75, 20k steps, 100 demos per task.

## Usage

```bash
# Install
git clone https://github.com/CCCalcifer/Skip-Policy.git && cd Skip-Policy
bash scripts/init.sh

# One-click eval (auto-downloads checkpoint + data)
bash scripts/eval.sh task=push_button

# Or manually download and eval
pip install huggingface_hub
python -c "from huggingface_hub import hf_hub_download; hf_hub_download('AYouZhuA/SkiP-RLBench', 'checkpoints/open_drawer.pt', local_dir='.')"
python scripts/eval.py task=open_drawer snapshot=checkpoints/open_drawer.pt
```

## Citation

```bibtex
@misc{dai2026skipskiprefineefficient,
      title={SkiP: When to Skip and When to Refine for Efficient Robot Manipulation},
      author={Mingtong Dai and Guanqi Peng and Yongjie Bai and Feng Yan and Chunjie Chen and Lingbo Liu and Liang Lin and Xinyu Wu},
      year={2026},
      eprint={2605.15536},
      archivePrefix={arXiv},
      primaryClass={cs.RO},
      url={https://arxiv.org/abs/2605.15536},
}
```