Embodied intelligence · 2026

PRIME W0

Physical Priors from World Models
for Embodied Manipulation

Huabin Peng^1,4, Fagang Jin^1,†, Yancong Wei^1,3, Tengfei Ma^1,4,*, Chuhang Liu^1,3, Chao Yu^2,3, Liqing Chen⁵

1 LUMOS SZ LAB TEAM · 2 LUMOS ROBOTICS · 3 THU · 4 CUHK · 5 NTU
† TEAM LEADER * CORRESPONDING AUTHOR

Paper PDF Watch real robots

Hierarchical VLM → ActionAsynchronous world-model conditioningCovariance-structured flow matching

02 / ABSTRACT

A policy with
physical memory.

VLA models are compelling instruction followers, but the physical world requires more than the current frame.

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.

03 / REAL ROBOT

Physical skills,
in increasing difficulty.

Five real-robot demonstrations grouped by the physical challenge each one exposes.

Choose a capability to focus the examples. Native video controls support scrubbing, playback speed, volume, and fullscreen.

Cloth folding with handoffsCHALLENGE 04 · FLEXIBLE + BIMANUAL

Dual-arm manipulationDEMO / 01

Manipulation in shared workspaceCHALLENGE 03 · HUMAN PROXIMITY

Person nearby during executionDEMO / 02

Small-object insertionCHALLENGE 02 · PRECISION

Pen into narrow containerDEMO / 03

Coordinated cloth flatteningCHALLENGE 04 · FLEXIBLE + BIMANUAL

Two arms control deformable fabricDEMO / 04

Sequential grasp and placementCHALLENGE 03 · MULTI-PHASE

Object transfer into target cupDEMO / 05

Challenge labels describe the visible manipulation demands. Use the timeline to scrub to any moment or enter fullscreen for a closer view.

04 / METHOD

From vision hierarchy
to action dynamics.

PRIME W0 preserves the VLM's rich, layer-wise knowledge while adding temporal physical priors precisely where action generation needs them.

OBSERVEVLM hierarchy

PREDICTWorld-model subgoal

ACTAdaptive MoE + ADM

01 / REPRESENTATION

Per-layer KV injection

Hidden states from a pretrained VLM are injected at every action layer, letting an Adaptive MoE expert select the representation depth the current action requires.

02 / DYNAMICS

World-model semantic bridge

An asynchronously predicted subgoal image is paired with a DINOv2 bridge over historical frames, so the controller remains grounded in how the scene is changing.

03 / CONTROL

Action Dynamic Manifold

Covariance-structured noise aligns flow matching with robot kinematic coupling, shaping a more physically coherent action distribution.

PRIME W0 model architecture — **Figure 1 · PRIME W0 architecture.** Hierarchical VLM features, asynchronous world-model subgoals, historical semantics, and the ADM action expert.

Four-stage training pipeline for PRIME W0 — **Figure 2 · Four-stage training recipe.** VLM pre-training, DAT/NTP visuomotor pre-training, DiT decoder assembly with world-model injection, and end-to-end SFT.

05 / RESULTS

Strong in simulation.
Stronger in the real world.

PRIME W0 pairs state-of-the-art LIBERO performance with consistent gains on contact-rich physical tasks.

TABLE 1 · LIBERO TASK SUCCESS (%)

Method	Spatial	Object	Goal	Long	Average
OpenVLA [15]	84.7	88.4	79.2	53.7	76.5
π0-FAST [25]	96.4	96.8	88.6	60.2	85.5
GR00T-N1 [24]	94.4	97.6	93.0	90.6	93.9
π0 [5]	96.8	98.8	95.8	85.2	94.2
UniVLA [7]	95.4	98.8	93.6	94.0	95.5
Discrete Diffusion VLA [19]	97.2	98.6	97.4	92.0	96.3
OpenVLA-OFT [16]	97.6	98.4	97.9	94.5	97.1
MemoryVLA [30]	98.4	98.4	96.4	93.4	96.7
π0.5 [26]	98.8	98.2	98.0	92.4	96.9
FLOWER [28]	97.5	99.1	96.1	94.9	96.9
EO-1 [27]	99.7	99.8	99.2	94.8	98.4
PRIME W0 (Ours)	98.8	100.0	99.0	98.6	99.1

PRIME W0 reaches 99.1% average success, with the strongest score on LIBERO-Object (100.0%) and LIBERO-Long (98.6%). Bold marks best; underline marks second best.

TABLE 2 · REAL-ROBOT TASK SUCCESS (%)

Task	OpenVLA	GR00T-N1.6	π0.5	PRIME W0
Pick-and-place	39.1	75.3	81.4	93.7
Stacking	25.2	61.8	67.9	70.1
Cloth folding	4.6	21.2	85.4	89.1
Small-object grasping	8.3	26.7	59.2	61.8
Assembly	1.4	22.1	51.8	75.9
Average	15.7	41.4	69.1	78.1

All methods use the same physical hardware under identical conditions. Gains are largest in temporally extended, contact-rich tasks: cloth folding (+3.7 points), small-object grasping (+2.6), and assembly (+24.1) over π0.5. These tasks require stable phase tracking across control cycles—the regime targeted by world-model subgoal injection, historical tokens W_hist, and phase-aware MoE routing.

06 / CONTRIBUTIONS

Three priors,
one policy.

HIERARCHY

VLM features without collapse

Layer-wise KV injection avoids discarding the detailed visual structure distributed throughout a pretrained VLM.

TEMPORALITY

World models at action time

Asynchronous subgoal prediction and historical-frame semantics make the control process explicitly dynamics-aware.

PHYSICS

Manifold-aware actions

ADM flow matching respects kinematic correlations instead of treating each action dimension as independent noise.

07 / PAPER

Read and cite.

The full paper describes the four-stage training recipe, benchmark results, and component ablations.

PAGE 01 · LIVE PDF

PREPRINT

PRIME W0:
Physical Priors from World Models

Live preview from the current PDF. Replace the PDF file to update this preview automatically.

Open full paper ↗

BIBTEX

@article{peng2026primew0,
  title={PRIME W0: Physical Priors from World
    Models for Embodied Manipulation},
  author={Peng, Huabin and Jin, Fagang and Wei,
    Yancong and Ma, Tengfei and Liu, Chuhang and
    Yu, Chao and Chen, Liqing},
  journal={arXiv preprint},
  year={2026}
}