Whole-body humanoid learning without teleoperation

WEVO

World Model-Enabled VLA Optimization for Whole-Body Loco-Manipulation

Qianzhong Chen1,2* Muyuan Lin1 Yuyin Sun1 Mac Schwager2 Cheng-Hao Kuo1 Roberto Martín-Martín1,3

1Amazon 2Stanford University 3The University of Texas at Austin

*Work done during an internship at Amazon.

0 target-task teleoperation
20 real rollouts per task
480 WM rollouts per task
75.0% mean video WM RL success

00

Abstract

Reinforcement learning can improve vision-language-action policies beyond imitation, but humanoid exploration is unsafe and simulator RL can exploit residual sim-to-real gaps. We present WEVO, a teleoperation-free framework for post-training a whole-body humanoid VLA in an action-conditioned video world model. Simulated manipulation and locomotion play data improve contact-dynamics and viewpoint-change prediction. A VLA trained on synthetic demonstrations collects 20 autonomous real robot rollouts per task to align the video WM with deployment observations. The VLA, whole-body controller, and video WM then form a closed-loop RL environment with dense progress rewards. Across three real robot tasks, video WM RL raises mean success from 23.3% to 75.0%, compared with 48.3% for real robot RL. The system leverages complementary environments: simulation provides policy initialization and broad coverage, while the video WM provides a deployment-aligned environment for policy improvement.

01

Sim to dream to real

Complementary roles for simulation and video WM

WEVO system pipeline
The LEGS simulator supplies synthetic task demonstrations and play data. Autonomous real robot rollouts align the video WM with deployment observations, after which the VLA, video WM, and SONIC–MuJoCo whole-body controller form a closed-loop policy-improvement environment with dense progress rewards.
1

Initialize in simulation

The LEGS simulator generates 200 whole-body demonstrations per task for base-VLA training and physically grounded play data for broad dynamics coverage.

2

Align the video WM

Each task uses 20 autonomous real robot rollouts, balanced between successes and failures, to align generated observations with deployment.

3

Improve the policy

The current VLA generates 480 on-policy video WM rollouts, while stage-aware progress supplies dense rewards without further robot interaction.

02

LEGS simulator

Task demonstrations and play data

Full-state planning supplies task data, while autonomous play data covers contact outcomes and locomotion-induced viewpoint changes.

TASK DATA

Synthetic demonstrations

Whole-body action sequences are generated from full-state planning and locomotion and manipulation primitives.

PLAY DATA

Manipulation

Randomized objects, grasp offsets, slips, misses, and collisions provide broad supervision for contact dynamics.

PLAY DATA

Locomotion

Approaches, turns, and long-range walking expose the model to large egocentric viewpoint changes.

03

Deployment-aligned video WM

Replaying real actions through generated observations

We replay held-out action trajectories and compare autoregressive video WM predictions with temporally aligned ground-truth observations.

Pick-and-Place Bottle. The generated rollout remains aligned with the recorded egocentric trajectory under the same actions.
PLAY DATA MID-TRAINING

Improves every fidelity metric

Under a matched 45,000-step budget, play data improves average PSNR by 0.93 dB and interaction-region PSNR by 1.17 dB. The larger interaction-region gain indicates stronger modeling around the robot hands and manipulated objects.

Pick-and-Place Bottle +1.24 PSNR +1.59 interaction PSNR
Pick-Turn-Place Bottle +0.70 PSNR +0.90 interaction PSNR
Bimanual Move Basket +0.85 PSNR +1.01 interaction PSNR

04

Closed-loop policy improvement

The VLA acts inside the video WM

Each video WM call receives one seed frame and achieved 43-DoF joint states, then generates 12 future views. The terminal generated frame seeds the next call, closing the recurrent VLA–WBC–video WM loop. A progress model converts each trajectory into dense rewards.

Pick-and-Place Bottle

Pick-Turn-Place Bottle

Bimanual Move Basket

Dense progress prediction over a video WM rollout
Stage-aware progress provides step-level credit, including on failed and suboptimal video WM rollouts.

05

Real robot evaluation

Video WM RL improves all three tasks

Video WM and real robot success rise together over 15 RL rounds. In contrast, direct simulator RL reduces real robot success on every task, showing that simulation is effective for initialization but not final policy improvement.

Autonomous real robot executions and DICE-RL policy improvement across the three tasks. Video WM success tracks real robot success more closely than LEGS simulator success.
Behavior cloning 23.3% mean success
Simulator RL 10.0% mean success
Real robot RL 48.3% mean success
Video WM RL 75.0% mean success
Method Pick-and-Place Bottle Pick-Turn-Place Bottle Bimanual Move Basket Mean
Baselines
Behavior cloning 40% 20% 10% 23.3%
Real robot RL 60% 45% 40% 48.3%
Simulator RL 30% 0% 0% 10.0%
VLAW 45% 20% 35% 33.3%
WMPO 55% 30% 35% 40.0%
RL algorithms in WEVO
AWR 75% 70% 70% 71.7%
ReinFlow 60% 50% 45% 51.7%
DICE-RL 85% 75% 65% 75.0%
Sparse reward45.0%
Fixed offline rollouts58.3%
Without play data63.3%
Full pipeline75.0%

06

How the two environments work together

Initialization

Synthetic demonstrations

Photorealistic simulation trains a whole-body base policy with enough hardware success to collect autonomous deployment experience.

Coverage

Play data

Physically grounded manipulation and locomotion experience improves contact dynamics and consistency under large viewpoint changes.

Alignment

Real robot rollouts

Limited autonomous successes and failures align the video WM with deployment observations and cross the remaining visual gap.

Improvement

Dense on-policy RL

Dense progress and on-policy training extract useful signal from successful, failed, and suboptimal video WM rollouts.

07

Citation

BibTeX

@article{chen2026wevo,
  title  = {WEVO: World Model-Enabled VLA Optimization
            for Whole-Body Loco-Manipulation},
  author = {Chen, Qianzhong and Lin, Muyuan and Sun, Yuyin and
            Schwager, Mac and Kuo, Cheng-Hao and
            Mart{\'i}n-Mart{\'i}n, Roberto},
  year   = {2026}
}