Robotics · 2026-10-30

PRIME W0: Physical Priors from World Models for Embodied Manipulation

Huabin Peng, Fagang Jin, Yancong Wei, Tengfei Ma, Chuhang Liu, Chao Yu, Liqing Chen

arXiv:2610.39528PDF

AI-assisted summary based on the paper. Consult the linked source for methods and limitations. Benchmark numbers appear after editor verification.

TL;DR

PRIME W0 is a hierarchical conditional flow-matching VLA for embodied manipulation. It injects per-layer VLM KV into an Adaptive MoE action expert, uses asynchronous world-model subgoal-image prediction plus a DINOv2 historical-frame semantic bridge for dynamics-aware conditioning, and generates actions with Action Dynamic Manifold (ADM) flow matching using covariance-structured noise that respects kinematic coupling.

Why it matters

The authors target two common VLA limitations—temporal myopia and semantic compression—by integrating explicit future observation priors from a world model into a VLM-action pipeline. Their approach aims to enable temporally coherent, instruction-grounded, and physically feasible continuous control via closed-loop inference and manifold-constrained action generation.

Method

  • Per-layer KV injection: extract Key/Value projections from all VLM backbone layers and inject them into an action expert via cross-attention.
  • Dynamics-aware conditioning: asynchronous world-model prediction of subgoal RGB images plus a DINOv2 historical-frame semantic bridge for long-horizon context.
  • Manifold-constrained control: Action Dynamic Manifold (ADM) flow matching over an Adaptive MoE decoder with covariance-structured noise respecting kinematic coupling.
Abstract (from arXiv)

Vision-Language-Action (VLA) models excel at instruction following, yet most remain feed-forward: they condition actions only on the current observation and discard both long-horizon temporal context and the hierarchical structure of VLM representations. We present PRIME W0, a hierarchical conditional flow-matching framework that closes this gap through three synergistic designs: (1)per-layer KV injection from a pretrained VLM into an Adaptive MoE action expert; (2)dynamics-aware conditioning via asynchronous world-model subgoal-image prediction and a DINOv2 historical-frame semantic bridge; and (3)Action Dynamic Manifold (ADM) flow matching with covariance-structured noise that respects robot kinematic coupling. PRIME W0 is trained in four stages---VLM pre-training, DAT/NTP visuomotor pre-training, DiT flow-matching decoder assembly with WM injection, and end-to-end SFT---and achieves state-of-the-art results on LIBERO and real-robot benchmarks, with systematic ablations validating each architectural component.

Related papers

Publication record

Bibliographic evidence describes where and when a record appeared. It does not validate scientific claims.

First public release: 2026-10-29 · Indexed by Pub-AI: 2026-10-30 UTC

preprint

PRIME W0: Physical Priors from World Models for Embodied Manipulation

Huabin Peng, Fagang Jin, Yancong Wei, Tengfei Ma, Chuhang Liu, Chao Yu, Liqing Chen

Preprint posted: 2026-10-29 · Original record ↗

arxiv
2610.39528
Explicit source versions (1)
  • v1 · 2026-10-29
  1. Preprint available

    Source date: 2026-10-29 · First detected: 2026-10-30 UTC · Historical baseline

    Source evidence (2)
Abstract · arxiv

Vision-Language-Action (VLA) models excel at instruction following, yet most remain feed-forward: they condition actions only on the current observation and discard both long-horizon temporal context and the hierarchical structure of VLM representations. We present PRIME W0, a hierarchical conditional flow-matching framework that closes this gap through three synergistic designs: (1)per-layer KV injection from a pretrained VLM into an Adaptive MoE action expert; (2)dynamics-aware conditioning via asynchronous world-model subgoal-image prediction and a DINOv2 historical-frame semantic bridge; and (3)Action Dynamic Manifold (ADM) flow matching with covariance-structured noise that respects robot kinematic coupling. PRIME W0 is trained in four stages---VLM pre-training, DAT/NTP visuomotor pre-training, DiT flow-matching decoder assembly with WM injection, and end-to-end SFT---and achieves state-of-the-art results on LIBERO and real-robot benchmarks, with systematic ablations validating each architectural component.