Atom-0 studies how egocentric human experience can improve robot policies during pre-training. We compare three ways to bridge the human–robot embodiment gap: domain-specific action heads, progressive embodiment alignment, and joint world–action modeling.
This code release includes Atom-DH, Atom-CL, and Atom-WAM, with robot-only baseline recipes. The WAM code has been adapted to the manuscript's 50×80, image-and-language-only interface; it has not been retrained. See the paper alignment audit for remaining corpus and reproducibility gaps.
| Route | Method | Code in this release |
|---|---|---|
| Atom-DH | Joint ego–robot co-training with a shared backbone and separate ego/robot action projections; optional trajectory-aware OT alignment | Code and training guide |
| Atom-CL | Ego pre-training → human–robot embodiment alignment → robot pre-training → Piper post-training | Training guide; implementation in the repository root |
| Atom-WAM | Fast-WAM-based joint future-video and action modeling | Code and training guide |
| Robot-only baselines | Robot-only pre-training followed by downstream adaptation | VLA and WAM baseline guide |
The VLA routes use the π₀.₅ architecture with base PaliGemma VLM initialization for the paper experiments. Loading a pretrained π₀.₅ policy is supported by the underlying code, but changes the initialization protocol.
The current manuscript reports that progressive alignment gives the strongest real-robot results. Both VLA routes improve OOD performance over the VLA robot-only baseline. Direct ego–robot WAM co-training underperforms the WAM robot-only baseline in this evaluation.
Seven tasks are evaluated with 10 ID and 10 OOD trials per task, giving 70 ID, 70 OOD, and 140 combined trials per checkpoint. Each OOD setting preserves the task goal and changes a spatial, appearance, or physical factor.
| Family | Method | ID success | OOD success | Combined success |
|---|---|---|---|---|
| VLA | Robot-only baseline | 42.86% (30/70) | 22.86% (16/70) | 32.86% (46/140) |
| VLA | Atom-DH | 44.29% (31/70) | 42.86% (30/70) | 43.57% (61/140) |
| VLA | Atom-CL | 60.00% (42/70) | 42.86% (30/70) | 51.43% (72/140) |
| WAM | Robot-only baseline | 28.57% (20/70) | 28.57% (20/70) | 28.57% (40/140) |
| WAM | Atom-WAM | 20.00% (14/70) | 21.43% (15/70) | 20.71% (29/140) |
These are manuscript-reported results for the best measured checkpoint from each route, synchronized on 2026-09-07. They are not results of the release validation checks or the newly modified WAM code. See evaluation protocol and representation analysis.
Atom-0/
├── routes/
│ ├── atom_dh/ # Independent OpenPI project: dual heads and OT
│ ├── atom_cl/ # Guide to the root progressive-training project
│ └── atom_wam/ # Independent PyTorch/OpenPI world–action project
├── baselines/ # Robot-only baseline recipes and release status
├── evaluation/ # VLA and robot-only baseline offline evaluators
├── src/openpi/ # Atom-CL models, data loading, training and policies
├── scripts/ # Atom-CL conversion, normalization, training and serving
├── packages/ # Policy client
├── assets/ # Small normalization statistics and action-space metadata
├── tests/ # Root-project tests
├── docs/ # Data contracts, evaluation and release notes
├── experiments/ # Atom-CL stages and auxiliary VLA experiments
└── third_party/ # Upstream integrations
Atom-DH, Atom-CL and Atom-WAM contain different versions of the openpi package. Use separate Python environments and run commands from the route's project directory. The root layout is retained so existing Atom-CL scripts continue to work. Supporting KI/MEM/DCC experiments are grouped under experiments/auxiliary_vla; they are not additional paper routes. Public navigation uses method names, while internal model/config identifiers remain compatible with existing checkpoints.
Use Linux, Python 3.11, and an NVIDIA GPU for model training. The projects pin JAX/Flax and other dependencies in their respective pyproject.toml and uv.lock files.
git clone --recurse-submodules https://github.com/Agentic-Intelligence-Lab/Atom-0.git
cd Atom-0
# Atom-CL / root project
GIT_LFS_SKIP_SMUDGE=1 uv sync --python 3.11 --group rldsFor Atom-DH, create its environment from its own directory:
cd routes/atom_dh
GIT_LFS_SKIP_SMUDGE=1 uv sync --python 3.11 --group rldsChoose one route guide before launching training. It specifies dataset roots, normalization assets, initialization weights, and the correct configuration names. Dataset conversion and full normalization must match the selected recipe.
The manuscript's curated recipe contains 334,054 episodes, approximately 2,659 hours, from robot demonstrations, EgoVerse, and task-matched human–robot alignment demonstrations. The initial pool is approximately 3,033 hours before filtering. The data guide describes the sources and input contract.
Both VLA routes use a semantic 80-dimensional state/action space, a 50-step action horizon, three canonical image slots, per-dataset normalization, and masks for unavailable action dimensions and views. Alignment-stage motion representations require the matching converter and statistics; they must not be interchanged with ordinary joint-space targets.
Raw datasets, videos, model weights, optimizer state, training logs, and credentials are not distributed in this repository. The JSON files under assets/ contain small normalization and schema metadata. Obtain pretrained weights from their original providers under the applicable terms.
- Atom-CL: four-stage execution of the progressive route
- Atom-DH: dual heads, OT ablation, and Piper fine-tuning
- Atom-WAM: paper-aligned world–action modeling
- Robot-only VLA and WAM baselines
- Remote policy serving
- Release provenance and validation scope
The paper describes three pre-training stages for Atom-CL. The repository additionally names downstream Piper fine-tuning Stage 4. Checkpoint directories are zero-indexed: for example, a completed 20,000-update run can end at 19999.
Atom-0 is developed at the Agentic Intelligence Lab and builds on Physical Intelligence's OpenPI. See upstream provenance, contributors, and CITATION.cff.
Code is distributed under Apache-2.0, subject to retained third-party notices. Gemma-related terms are retained in LICENSE_GEMMA.txt; model and dataset licenses remain separate from the code license. A public paper link will be added when available.