ZJU's VLA-Rolling reuses discarded action tails to cut VLA inference cost across 16 model-benchmark pairs
OmniAI-ZJU · hf · 2026-10-21
VLA-Rolling from Zhejiang University's OmniAI Group targets the receding-horizon paradigm of flow-matching VLA models, where only a short prefix of each action chunk is executed and the remaining "action tail" is discarded. The paper shows this tail retains transferable denoising progress: tail initialization can match or beat a native 10-step Gaussian baseline with fewer denoising steps.
- Three stages: action-tail inheritance, per-action confidence estimation, and confidence-guided refinement via renoising/denoising budget allocation, with the underlying VLA frozen
- Optimal denoising budgets vary per sample and should adapt to updated observations
- Across SmolVLA, X-VLA, MolmoAct2 on LIBERO, MetaWorld, RoboTwin 2.0 — 16 model-benchmark pairs — it cuts action-chunk inference cost and boosts success rates
More from Embodied
- Video-Point Model learns 4D world models via 3D point trajectory forecasting — Byungwoo Jeon · 2026-10-21
- Japan deploys slightly terrifying giant robots for railway maintenance — CurieuxExplorer · 2026-10-21
- Next Xbox Reportedly to Cost $1,200-$1,800 in 2028 as AI Devours Memory Supply — aakashgupta · 2026-10-21
- Maker 3D-prints an adapter to mount a chuck onto a robot motor — _Stocko_ · 2026-10-21
- Phonon TTS pipeline ported to GenieX, targeting on-device NPU inference — HowDevelop · 2026-10-21
- Ropedia releases open W0 world model trained on 10,000 hours of egocentric human video — liuziwei7 · 2026-10-21