Identity and contribution
Heng Fang, Shangru Li, Shuhan Wang, and colleagues introduce DOMINO, a dataset/benchmark, and PUMA, a dynamic-manipulation policy. The notes use arXiv:2603.15620v3, June 30, 2026, a 43-page manuscript; the project records ECCV 2026. DOMINO is distinct from DynamicVLA’s DOM.
The work studies motion-sensitive representations and dynamic data. It does not introduce RTC-style joining, FLASH verification, or LAAS prefix removal. The appendix leaves real control-delay integration in the simulator benchmark for future work, limiting direct comparisons with latency-aware systems.
Dataset and dynamic assumptions
DOMINO contains 117,000 expert trajectories, 35 tasks, and five robot configurations. Main experiments use Aloha-AgileX, clean scenes, Level 1, DOMINO@0.1, rather than jointly validating all embodiments and randomized conditions.
Level 1 uses planar constant-velocity motion; Level 2 uses polynomial curves; Level 3 joins two or three independently sampled segments with abrupt transitions. The body and appendix differ in sampling details. In particular, some Level 2 acceptance checks constrain average speed rather than every instantaneous speed.
SAPIEN/RoboTwin generation first finds a static feasible robot path, then places the moving object so it reaches the desired contact location at the correct time:
Before grasp contact, the target follows a prescribed kinematic path. On contact it becomes a physics-controlled body and stops its autonomous motion. This retains post-contact physics but excludes continued external driving and some coupled dynamics.
PUMA inputs and auxiliary supervision
The backbone is Qwen3-VL-4B. The actual configuration uses language, current multiview images, and historical flow, without extra proprioception. A parallel MLP predicts fifteen-step chunks of fourteen-dimensional absolute joint actions.
Four historical frames spaced four frames apart are compared with the current frame. Farneback flow on 64×64 grayscale images is color-coded, percentile-normalized, and thresholded before visual encoding. Flow includes image-plane motion from objects, cameras, and occlusion; it is not separated metric 3D velocity. Training caches flow; deployment computes it from a history buffer.
World queries predict future object features. Frozen GroundingDINO and SAM2 locate/mask the task object; frozen DINOv2-B/14 provides pooled future-region targets. Default supervision covers four future frames spaced four frames apart. Object parsing and segmentation errors affect the targets.
The combined objective is:
This is imitation plus auxiliary feature prediction, not an implemented RL cost-minimization algorithm. Future target frames are training supervision, not inference inputs. Feature similarity does not directly enforce metric object trajectories or calibrated future uncertainty.
Training uses eight A100s, full-parameter fine-tuning, mixed precision, checkpointing, and ZeRO-2. Real adaptation uses LoRA from simulation weights.
Evidence and metrics
Simulation evaluates 100 episodes per task. SR requires complete task success. Manipulation Score (MS) combines final end-effector/object geometry with penalties; it is neither success probability nor a full contact-quality measure. The displayed route formula omits explicit negative clipping and zero-denominator protection; implementation needs checking. Taking the better arm’s progress can underweight the other arm’s failure.
| Main setting | SR | MS |
|---|---|---|
| OpenVLA | 1.54% | 6.10 |
| RDT-1B | 5.34% | 17.71 |
| π₀ | 8.17% | 23.96 |
| π₀.₅ | 9.63% | 26.17 |
| OpenVLA-OFT, Prismatic | 9.06% | 24.06 |
| OpenVLA-OFT, reimplemented Qwen3-VL | 10.86% | 30.49 |
| PUMA | 17.20% | 34.97 |
The 6.34-point gain over the same-backbone OFT is not a 6.34% relative increase. About 82.8% of episodes still fail. Static-to-dynamic drops are large; simply adding dynamic data gains less than three points for several baselines. Giving OFT privileged future trajectories improves 9.06%→10.33% SR and 24.06→32.00 MS: spatial guidance alone does not solve contact.
Flow without future supervision gives 11.71%; two future targets 14.80%; four 17.20%; six 13.62%. Raw RGB history plus two targets gives only 8.15% in this configuration. Longer prediction is not always better. Interception is 24.2% while sustained tracking is 8.9%.
Adapted Level 2/3 results are 10.5%/4.6%, after target-level LoRA, not zero-shot transfer. Mixed-data and doubled-data results use different prediction counts or task subsets and should not be merged with the main table.
Real tasks, latency, and project use
Five real Piper tasks use fifty demonstrations and twenty evaluations each. PUMA successes are 50%, 45%, 35%, 25%, and 55%, averaging 42%, versus π₀.₅ 24%. Controlled string-driven straight motion approximates Level 1; arbitrary curves, collisions, and unseen scenes are not established.
On an RTX 4090, mean inference is 103.7 ms, including 98.3 ms backbone, 3.45 ms flow, and 2.0 ms preprocessing, about 9.6 queries/s. Multiplying by fifteen chunk steps does not give a fresh-vision feedback rate.
Independent analysis: PUMA improves motion representation; scheduling/latency methods solve a different interface. Dynamic data, successful contact, and timely real-world response remain separate targets. Bind tasks, embodiment, clean/randomized setup, levels, speed interpretation, sampling gaps, action normalization, segmentation caches, stop-on-contact rules, and MS code. Add real elapsed simulator motion and observation-age reporting when studying execution latency. The official repository’s simulator/policy environments and episode manifests aid reproducibility; the source notes did not run them.
Source references
The source versions and checks above are retained from the uploaded PaperRead notes. This English edition edits their presentation; it does not report new experiments.