CoRL paper: one-step inference matches multi-step for large VLAs, 1.6-4.8x faster action generation
kaixhin · x · 2026-10-26
Araya's RL team will present "When One Step Is Enough" at CoRL 2026, empirically testing whether flow-matching robot policies really need multi-step Euler integration at inference.
- Key finding: for large, well-trained VLAs, a single Euler step largely matches multi-step task success, with 1.6x-4.8x faster action generation — no retraining, distillation, or auxiliary objectives needed.
- Coverage: SmolVLA (450M), X-VLA (900M), FLOWER (947M), π₀.₅ (3.3B), LAP (3.3B), XR-0 (4.7B) plus a 211M DiT baseline, evaluated on LIBERO/LIBERO+, CALVIN ABC→D, RoboTwin 2.0, a real-world Franka dataset, and D3IL Avoiding.
- Scaling pattern: equivalence strengthens with model size, VLM pretraining, and fine-tuning; smaller VLAs deviate more on long-horizon tasks (though k-step wasn't consistently better).
- Behavioral difference: 1-step inference reduces trajectory diversity entropy, concentrating on fewer modes; trajectory smoothness showed no significant difference.
Paper, dataset, and project site are publicly available.
More from Embodied
- Visual map of 12 humanoid robot makers: valuations, flagships and production targets — CurieuxExplorer · 2026-10-26
- Matt Wolfe: robot butler demo and this week's AI news roundup — Matt Wolfe · 2026-10-26
- Kimi K3.1 Pro scores 21.3% on humanoid BiGym 2 benchmark at 20x GPT-6 Sol's cost — teortaxesTex · 2026-10-26
- Reverse-engineered Ray-Ban Meta glasses show chess moves in AR at 3000+ Elo — Vjeux · 2026-10-26
- HKU/ZJU/Fudan with Lightspeed launch HiSTEVE: hierarchical VLA agent playing Minecraft via native mouse and keyboard — tencent · 2026-10-26
- Hobbyist teaches StackChan robot to read scripts with Grok and Hermes-Agent — alexcovo_eth · 2026-10-26