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

[Paper] [Code]

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. Rather than conditioning a universal policy on language, DISC uses a hypernetwork to generate the entire parameter set of a task-specific visuomotor policy from the instruction alone. This repository provides training, evaluation, and test-time training code for reproducing the results reported in the paper.

📦 Checkpoints

The repository currently contains two checkpoints to reproduce the results reported in the paper:

• DISC-libero90.ckpt: Model trained on the LIBERO-90 dataset (Table 1).
• DISC-libero_cotrain_spatial_object_goal.ckpt: Model co-trained on LIBERO-spatial, LIBERO-object, and LIBERO-goal (Table 2).

📊 Evaluation

To evaluate a trained checkpoint on LIBERO-90, you can use the following command from the official repository:

python train.py --config-name=libero_90_eval model=hylap +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}
}