Checkpoints on RLBench-10 tasks are available on HuggingFace.

Quick Start

Environment Setup

conda create -n skip python=3.9 -y
conda activate skip
bash scripts/init.sh
source ~/.bashrc

This installs CoppeliaSim, RLBench, and all dependencies. See init.sh for details.

One-click Evaluation

Automatically downloads the pretrained checkpoint and evaluation data from HuggingFace, then runs evaluation:

bash scripts/eval.sh task=push_button

Training

# SkiP (default: W=16, q=0.75)
CUDA_VISIBLE_DEVICES=0 xvfb-run -a python scripts/train.py task=open_drawer kf_mode=skip

# SkiP with custom hyperparameters
python scripts/train.py task=open_drawer kf_mode=skip skip.window=16 skip.high_quantile=0.75

# CoA-rev baseline
python scripts/train.py task=open_drawer

# CoA-fwd baseline
python scripts/train.py task=open_drawer method.action_order=FORWARD

Evaluation

python scripts/eval.py task=open_drawer snapshot=path/to/checkpoint.pt

Evaluation outputs per-episode success/failure to the terminal and saves metrics to eval_metrics.json in the experiment directory:

{
  "summary": {
    "success_rate": {"mean": 0.72, "std": 0.0},
    "avg_steps_success": {"mean": 49.1, "std": 0.0},
    "avg_steps_all": {"mean": 85.8, "std": 0.0}
  }
}

To save evaluation videos, add log_eval_video=true.

Download RLBench Dataset

# Download all tasks
python scripts/download_dataset.py

# Download a specific task
python scripts/download_dataset.py --task reach_target --train-episodes 100 --eval-episodes 25

Configuration

All settings are in src/cfgs/launch.yaml. Key parameters:

Parameter	Default	Description
task	—	RLBench task name (required)
kf_mode	full	full (CoA), skip (SkiP), heuristic, cache
skip.window	16	ST-DCT window size W
skip.high_quantile	0.75	Quantile threshold q
num_train_steps	20000	Total training steps
batch_size	128	Training batch size
demos	100	Number of demonstrations
eval_every_steps	100000	Evaluation interval

Override via command line: python scripts/train.py task=push_button batch_size=64 kf_mode=skip

Directory Structure

scripts/          Training, evaluation, data downloading
src/
  cfgs/           Hydra configuration files
  dataset/        Dataset loading and SkiP relabeling
  envs/           RLBench environment and MSK segmentation
  methods/
    coa/          CoA backbone (shared by SkiP and baselines)
    act/          ACT baseline
    dp/           Diffusion Policy baseline
  workspace.py    Training workflow

Citation

@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}, 
}

Acknowledgement

This codebase is built upon Chain-of-Action. We thank the authors for their open-source implementation.

License

This project is licensed under the Apache License, Version 2.0. See the LICENSE file for details.