SAIL: Faster-than-Demonstration Execution of Imitation Learning Policies

Project page · Paper (arXiv:2506.11948)

This repository is the companion code for SAIL (Speed Adaptation for Imitation Learning): a full-stack system for executing visuomotor imitation-learning policies faster than demonstration speed while preserving task success.

It is built on robomimic and adds dataset processing, training configs, Error-Adaptive Guidance (EAG) evaluation, and adaptive speed modulation used in the paper.

Release timeline

Phase	What	Status
1	Simulation — training, dataset processing, EAG evaluation, and the robosuite SAIL patch
2	Real hardware — Franka (Deoxys) and UR5 (ROS 2) stacks from the paper

Installation

Requires conda. From the repository root:

bash robomimic/SAIL/installation.sh
conda activate SAIL

The install script will:

1. Create/reuse a conda env (SAIL by default) with PyTorch 2.1
2. Clone ARISE robosuite at a pinned v1.4.1 commit and apply third_party/patches/robosuite-v1.4.1-sail.patch
3. Install this package editable
4. Install stock AWE for waypoint extraction

Important: apply the robosuite patch only against the pinned commit (b9d8d3de). See third_party/patches/README.md.

Environment overrides:

Variable	Meaning
SAIL_ENV_NAME	Conda env name (default SAIL)
SAIL_PYTORCH_CUDA	CUDA build for conda pytorch (default 11.8)
ROBOSUITE_INSTALL_DIR	Where to clone/patch robosuite
AWE_INSTALL_DIR	Where to clone AWE

Dataset processing

Create reached-pose actions and adaptive-speed labels:

# Reached actions
python robomimic/SAIL/dataset_processing/add_all_actions.py --dataset=<PATH_TO_DATASET>

# AWE waypoints + precision / speed labels
python robomimic/SAIL/precision_processing/save_awe_waypoint_concurrent.py --dataset=<PATH_TO_DATASET>
python robomimic/SAIL/precision_processing/label_awe_trajectory_precision.py --dataset=<PATH_TO_DATASET>

Tips:

• Tune --err_threshold in save_awe_waypoint_concurrent.py to match waypoint density from the AWE paper.
• Tune clustering parameters in label_awe_trajectory_precision.py for precise vs non-precise balance.
• Defaults work well for robomimic Lift / Can / Square.
• For repetitive pick-and-place demos, label sub-tasks separately if needed.

Visualize speed labels (control-frequency graph + agentview overlay):

python robomimic/SAIL/precision_processing/dynamic_speed_awe.py \
  --dataset=<PATH_TO_DATASET> --video_path=<OUTPUT.mp4>

Training

Example Diffusion Policy config with EAG conditioning:

robomimic/exps/templates/diffusion_policy_SAIL.json

python robomimic/scripts/train.py --config=robomimic/exps/templates/diffusion_policy_SAIL.json

Edit the config train.data path to point at your processed dataset.

Evaluation (receding horizon + EAG)

python robomimic/SAIL/run_trained_agent_receding_horizon.py \
  --agent=<PATH_TO_CHECKPOINT.pth> \
  --guide_config=robomimic/SAIL/guide_template/base_cfg_weight_1.json

Example guide configs are under robomimic/SAIL/guide_template/. Controller JSONs for high-fidelity tracking are under robomimic/robosuite_configs/*_SAIL.json.

Citation

If you use this code, please cite SAIL:

@article{arachchige2025sail,
  title={SAIL: Faster-than-Demonstration Execution of Imitation Learning Policies},
  author={Arachchige, Nadun Ranawaka and Chen, Zhenyang and Jung, Wonsuhk and Shin, Woo Chul and Bansal, Rohan and Barroso, Pierre and He, Yu Hang and Lin, Yingyan Celine and Joffe, Benjamin and Kousik, Shreyas and Xu, Danfei},
  journal={arXiv preprint arXiv:2506.11948},
  year={2025}
}

This repository extends robomimic; please also cite:

@inproceedings{robomimic2021,
  title={What Matters in Learning from Offline Human Demonstrations for Robot Manipulation},
  author={Ajay Mandlekar and Danfei Xu and Josiah Wong and Soroush Nasiriany and Chen Wang and Rohun Kulkarni and Li Fei-Fei and Silvio Savarese and Yuke Zhu and Roberto Mart\'{i}n-Mart\'{i}n},
  booktitle={Conference on Robot Learning (CoRL)},
  year={2021}
}

See CITATION.cff and NOTICE for third-party attributions (robosuite, AWE, Diffusion Policy, etc.).

Acknowledgments

This work was supported by the State of Georgia and the Agricultural Technology Research Program at Georgia Tech; the AI Manufacturing Pilot Facility / Georgia AIM (U.S. Department of Commerce EDA Award 04-79-07808); NSF CCF Award 2211815; and NSF Award 1937592.

We thank the ARISE Initiative for robomimic and robosuite, and the AWE authors for Automatic Waypoint Extraction.

License

MIT — see LICENSE. Upstream robomimic remains copyright Stanford Vision and Learning Lab; SAIL modifications are copyright Georgia Institute of Technology / the SAIL authors.