arXiv ↗
Code ↗
Weights ↗Atomic-motion annotation
- 01 Segment the motionUse robot movement and gripper changes to define physical motion segments.
- 02 Ground across viewsRead the scene through synchronized left, right, and wrist observations.
- 03 Describe & assessAssign atomic motions, task relevance, execution success, and task completion.
- 04 Check the annotationsReview the resulting labels with a checker before accepting the annotation.
Loading annotated demonstrations…
Execution-grounded rewards
With the WAM fixed, EG-GRPO trains the reasoner using rewards for motion appropriateness and execution success.
Reward definition
rtask · Is this the right next step?1 if the proposed motion is a suitable next step toward the task from the current state; 0 otherwise.
rexec · Did the robot execute it?1 if the rollout successfully completes the specified motion; 0 otherwise.
Paper ablation · RoboLab-120, default instructions, three runs: supervised reasoner 36.7% → GRPO without execution reward 34.2% → EG-GRPO 39.2%.
Atomic-WAM on RoboLab-120
The RoboLab organizers have reproduced Atomic-WAM using our released weights.
View official leaderboard ↗Vague instructions
| Policy | SR | Score |
|---|---|---|
| Atomic-WAM | 31.6% | 48.6 |
| BiMind v0.1 | 30.2% | 44.9 |
| VoLo | 29.5% | 45.4 |
| Cosmos3-Nano-Policy | 20.6% | 33.2 |
| Cosmos3-Edge-Policy | 15.4% | — |
| π0.5 | 15.2% | 28.1 |
| DreamZero | 14.9% | 27.0 |
| π0-FAST | 9.2% | 16.8 |
| GR00T N1.6 | 5.4% | 14.0 |
| paligemma-binning | 3.1% | 8.6 |
| π0 | 2.8% | 8.5 |
Default instructions
| Policy | SR | Score |
|---|---|---|
| Lakeside | 51.0% | 65.3 |
| FLUX 3 Action | 42.9% | — |
| HiDream-O1-Embodied | 39.9% | 50.5 |
| Atomic-WAM | 39.6% | 55.5 |
| OASIS WAM | 39.0% | 53.7 |
| Cosmos3-Nano-Policy | 36.8% | 51.9 |
| Phoenix | 34.4% | 45.9 |
| BiMind v0.1 | 33.3% | 47.2 |
| VoLo | 28.2% | 44.1 |
| π0.5 | 28.0% | 43.4 |
| DreamZero | 25.7% | 39.8 |
RoboLab-120 · October 13, 2026 snapshot
Due to version differences, the official leaderboard computes scores slightly differently from our paper.
Real-world demonstrations
We train on DROID and deploy on our Franka setup without further fine-tuning.
Our setup uses a RealSense D435 wrist camera instead of DROID’s ZED Mini, with different camera placements and viewpoints.
Put the blue pens into the bowl.
Wipe the spill, then put the red block into the cup.
Put the yellow objects into the cup.
Place the toy fox to the right of the bin.
Put all blocks in the right bin and all batteries in the left bowl.
Our current real-world setup uses synchronous inference. In the videos, pauses are sped up 8× and robot motion is sped up 2×.
Citation
@article{shi2026atomicwam,
title={Atomic-WAM: Execution-Grounded Atomic-Motion Reasoning for World Action Models},
author={Shi, Shijun and Liu, Zhe and Hou, Zhi and Chen, Yinan and Zhang, Tao and Hu, Kai and Liu, Yong and Zhang, Jiangning},
journal={arXiv preprint arXiv:2610.20157},
year={2026}
}