This repository provides simulation training and real-robot deployment for PRISM, based on HoloSoma.
| Branch | Use |
|---|---|
main |
Simulation: teacher training, rollout collection and student distillation |
sim2real |
Real-robot deployment with FastFoundationStereo |
For real-robot deployment, switch an existing clone to sim2real and follow that branch's README:
git fetch origin
git switch sim2real
git submodule update --init --recursiveLinux, Python 3.11 and a compatible NVIDIA GPU are required. See system requirements and environment details.
git clone --branch main https://github.com/amazon-far/PRISM.git
cd PRISM
python3.11 -m venv .venv
source .venv/bin/activate
bash install.sh
wandb loginDownload and prepare the Hugging Face dataset under data/:
bash download_data.shIncludes object meshes and prepares the student training shards. Dataset details.
Both training scripts use all visible GPUs on this machine. Set
CUDA_VISIBLE_DEVICES=0 for one GPU or CUDA_VISIBLE_DEVICES=0,1 for a subset.
Use --envs-per-gpu to lower memory use. Training details.
Train a privileged motion-tracking policy with PPO. Teacher-training data is still under review; supply a prepared teacher bank.
bash train_teacher.sh --motion-bank /path/to/teacher_bank --entity YOUR_WANDB_ENTITYCollect teacher trajectories and contact sidecars into outputs/rollout/.
bash rollout.sh --motion-bank /path/to/teacher_bankDistill the teacher into a depth policy using the prepared data under data/.
bash train_student.sh --entity YOUR_WANDB_ENTITYIf you use PRISM in your research, please cite:
@article{wang2026counterfactual,
title={Counterfactual Video Generation Enables Scalable Humanoid Loco-Manipulation},
author={Wang, Zihan and Wu, Zhen and Abbeel, Pieter and Duan, Rocky and Malik, Jitendra and Sferrazza, Carmelo and Liu, C. Karen and Shi, Guanya and Kanazawa, Angjoo},
journal={arXiv preprint arXiv:2609.38172},
year={2026}
}See LICENSE, NOTICE and THIRD_PARTY_LICENSES.
See CONTRIBUTING for security reporting.