PRISM loco-manipulation
SoftwareSimulation teacher, rollout and student-distillation code for humanoid loco-manipulation with a released 129-trajectory subset. [1] [3] [6]
Installation & first result →Official websiteSource codeGetting started
- For
- Humanoid robot learning researchers with NVIDIA simulation infrastructure. [1]
- Hardware
- Unitree G1 real-robot branch reported by authors [1]
- Source rights
- Some licenses are unresolved. Review licenses
- Unknowns
- operating system not fully established. Review gaps
Source researched · Not tested by us
Sources checked · Research details
Licenses & related hardware
- License
- Apache-2.0, Apache-2.0 card declaration; third-party assets separately scoped · license not fully assessed for model weights · Details by part
Overview
PRISM is a CoRL 2026 humanoid loco-manipulation project built on HoloSoma. Its main branch releases simulation teacher training, rollout and student-distillation code; a separate sim2real branch documents real-robot deployment. The released dataset contains 129 successful teacher-rollout trajectories, commands, contact sidecars, robot assets and checkpoints. The paper's real-robot behavior remains an author-reported result, not a result reproduced by this catalog. [1] [3] [6]
Is this for you?
- Workflows
- Prepare released data
- Train/distill student
- Evaluate in simulation; inspect separate deployment branch
What you need
- Languages / runtimes
- Python
- Dependencies
- HoloSoma
- NVIDIA GPU
- PRISM dataset
- Host computer
Linux, Python 3.11 and compatible NVIDIA GPU; multi-GPU training described.
- Compute for the workload
- Compatible NVIDIA GPU; scripts can use all visible GPUs, adjustable with CUDA_VISIBLE_DEVICES and environment count.
Installation & first result
Open the upstream getting-started guide [1]
- Install the pinned Linux/Python 3.11/NVIDIA environment.
- Download the checked dataset revision and prepare student shards.
- Use the documented student-distillation route; a prepared bank is needed for teacher retraining.
- Installation
install.sh in Linux Python 3.11 virtual environment; dataset downloaded separately.
Hardware, models & data
- Supported hardware
- Unitree G1 real-robot branch reported by authors; no compatibility test here.
- Inputs and outputs
Prepared motion/contact data and simulated observations to a depth-based student policy.
- File formats
- Motion NPZ
- Contact sidecars
- URDF/mesh assets
- PyTorch checkpoints
- Algorithms and models
- Privileged teacher PPO
- Depth-policy student distillation
- Evaluation
Simulation rollout and author-reported G1 demonstration; neither reproduced here.
Project notes
What can be attempted now
The documented setup requires Linux, Python 3.11, a CUDA 12.8 compatible NVIDIA driver, at least 30 GB for the environment and caches, and the separately downloaded dataset. The original recipes used 48 GB GPUs; that is a reported configuration, not a proven minimum. Lowering --envs-per-gpu can reduce memory demand but changes the global batch and may affect convergence. The download script pins and checks the dataset revision before preparing student shards. [1] [3] [4] [5]
The student-distillation path uses the released data and the repository's _ckpts/ checkpoints; it needs no new rollout. The three checkpoint files were present in the pinned repository archive and their SHA-256 values matched its manifest, but they were not loaded or run. The older checkpoints inside the dataset archive do not meet the current launcher's contract. Retraining the privileged teacher or collecting new rollouts requires a prepared teacher motion bank. [3] [5] [7] [8]
The raw videos and teacher-training motion bank are still under review, so the published materials do not establish complete from-video reproduction. The 129 trajectories are a filtered subset of 137 generated clips. Their packaged visual meshes are simplified; the dataset documentation says depth-observation equivalence to the original training meshes has not been validated. Original visual meshes are separately available, but that does not establish reproduction either. Dataset and source code declare Apache-2.0, while checkpoint, object and third-party rights need artifact-level checks. No data archive, training run, simulation rollout or G1 deployment was executed here. [1] [2] [3] [6]
What we checked
Source research only. No commercial relationship or hands-on testing. No upstream media copied to the public site.
Not yet established
- Requirements: operating system
Updates
- Upstream last updated
- Not established
- Latest release
- Not established
- Profile updated
- Nov 7, 2026
- Sources last checked
- Nov 7, 2026
- Next review
- Dec 7, 2026
- Reviewed revision
- Commit snapshot 0173e78fc62e3128dfb9edcf3706432405a5853e; no matching GitHub release found
Source files and licenses
Useful released student training and simulation code/data; teacher-training source inputs and raw videos are incomplete, and deployment is branch-specific.
| Layer | Availability / license |
|---|---|
| software | available · Apache-2.0 Artifact and declaration inspected at the cited scope. External assets/dependencies and actual operation need separate verification. [1] [2] |
| model weights | available · License not fully assessed Artifact and declaration inspected at the cited scope. External assets/dependencies and actual operation need separate verification. [7] |
| datasets | available · Apache-2.0 card declaration; third-party assets separately scoped Artifact and declaration inspected at the cited scope. External assets/dependencies and actual operation need separate verification. [6] |
| documentation | available · Apache-2.0 Artifact and declaration inspected at the cited scope. External assets/dependencies and actual operation need separate verification. [1] [2] |
References
Each source links to the exact revision we read. “Used for” shows where this page relies on it.
- S001 PRISM code and branches
Revision, date & where cited
amazon-far/PRISM contributors · github.com · revision
0173e78· README.md at · Accessed Nov 7, 2026
Used for: Overview, Is this for you?, What you need, Installation & first result and 3 more - S002 Apache-2.0 code license
Revision, date & where cited
amazon-far/PRISM contributors · github.com · revision
0173e78· LICENSE at · Accessed Nov 7, 2026
Used for: What can be attempted now, Source files and licenses - S003 Dataset preparation and limitations
Revision, date & where cited
amazon-far/PRISM contributors · github.com · revision
0173e78· docs/data.md at · Accessed Nov 7, 2026
Used for: Overview, Is this for you?, Installation & first result, What can be attempted now - S004 Installation requirements
Revision, date & where cited
amazon-far/PRISM contributors · github.com · revision
0173e78· docs/installation.md at · Accessed Nov 7, 2026
Used for: Installation & first result, What can be attempted now - S005 Training stages
Revision, date & where cited
amazon-far/PRISM contributors · github.com · revision
0173e78· docs/training.md at · Accessed Nov 7, 2026
Used for: Installation & first result, What can be attempted now - S006 PRISM dataset card
Revision, date & where cited
Amazon FAR · huggingface.co · Dataset card at · Accessed Nov 7, 2026
Used for: Overview, Is this for you?, What can be attempted now, Source files and licenses - S007 Repository checkpoint roles and limits
Revision, date & where cited
amazon-far/PRISM contributors · github.com · revision
0173e78· _ckpts/README.md at · Accessed Nov 7, 2026
Used for: What can be attempted now, Source files and licenses - S008 Repository checkpoint checksums
Revision, date & where cited
amazon-far/PRISM contributors · github.com · revision
0173e78· _ckpts/manifest.json at · Accessed Nov 7, 2026
Used for: What can be attempted now
Correction guidance · Discussion is not enabled · Profile format 0.7.0-draft.5