DISC: Decoupling Instruction from State-Conditioned Control via Policy Generation

Official implementation of the RSS 2026 paper:

DISC: Decoupling Instruction from State-Conditioned Control via Policy Generation

Hanxiang Ren, Pei Zhou, Xunzhe Zhou, and Yanchao Yang
Robotics: Science and Systems (RSS), 2026

🔍 Overview

DISC is a policy-generation framework that decouples high-level language instructions from state-conditioned low-level control. This repository provides training, evaluation, and test-time training code for reproducing the results reported in the paper.

🧩 Supported Task Suites

• LIBERO
  • libero_90
  • libero_spatial
  • libero_goal
  • libero_object
  • libero_10
• Meta-World: work in progress

⚙️ Installation

Clone and install the modified LIBERO dependency:

mkdir -p third_party
cd third_party
git clone https://github.com/Lifelong-Robot-Learning/LIBERO.git modified_libero
cd ..

uv sync
touch ./third_party/modified_libero/libero/__init__.py
uv pip install -e ./third_party/modified_libero

bash utils/post_install.sh

📦 Data Preparation

We use the OpenVLA version of LIBERO in our experiments. This version uses 256 x 256 image observations and filters out no-op actions.

Please download the official LIBERO dataset from the LIBERO repository, then convert it to the OpenVLA format using regenerate_libero_dataset.py from the OpenVLA repository.

After downloading and converting the datasets, organize them as follows:

/data/libero/
├── libero_90/
├── libero_object/
├── libero_spatial/
├── libero_10/
└── libero_goal/

🚀 Training

Train DISC on LIBERO-90:

python train.py --config-name=hylap

Train on other LIBERO suites:

python train.py --config-name=hylap data=libero_spatial
python train.py --config-name=hylap data=libero_goal
python train.py --config-name=hylap data=libero_object
python train.py --config-name=hylap data=libero_10

📊 Evaluation

Our checkpoints can be downloaded from DISC Huggingface Repo.

Evaluate a trained checkpoint on LIBERO-90:

python train.py --config-name=libero_90_eval model=hylap +ckpt_path=/path/to/checkpoint.ckpt

🔁 Test-Time Training

Run test-time training (TTT) on LIBERO-90:

python train.py --config-name=libero_90_ttt +ckpt_path=/path/to/checkpoint.ckpt

📝 Citation

If you find this repository useful for your research, please consider citing our paper:

@inproceedings{ren2026disc,
  title={{DISC}: Decoupling Instruction from State-Conditioned Control via Policy Generation},
  author={Hanxiang Ren and Pei Zhou and Xunzhe Zhou and Yanchao Yang},
  booktitle={Robotics: Science and Systems},
  year={2026},
  url={https://openreview.net/forum?id=i9ynkagJuj}
}

🙏 Acknowledgements

This codebase builds on resources from the LIBERO, OpenVLA, and HyPoGen projects. We thank the authors for releasing their datasets, environments, code, and tools.