WEVO: World Model-Enabled VLA Optimization for Whole-Body Loco-Manipulation
TL;DR WEVO post-trains a whole-body Unitree G1 VLA with RL inside an action-conditioned video world model, adapted with simulated play data and 20 autonomous real rollouts per task, with no teleoperation. Across three real loco-manipulation tasks, mean success rises from 23.3% to 75.0%, versus 48.3% for RL on the real robot.
What they do
Teleoperating a humanoid for VLA demonstrations is hard, RL on the hardware is unsafe, and RL in a simulator can exploit the remaining sim-to-real gap. WEVO gives the two environments separate jobs. A photorealistic Gaussian-splat simulator (LEGS) supplies demonstrations and play data. An action-conditioned video world model, adapted to a few autonomous real rollouts, then serves as the environment for RL post-training of a Unitree G1 whole-body VLA.
How it works
- Initialization. 200 synthetic episodes per task fine-tune a GR00T N1.7 VLA that outputs hand poses, gripper openings, base velocity and height for the SONIC whole-body controller. It then runs autonomously until 10 successes and 10 failures are collected per task.
- World model. The DreamDojo video model is mid-trained on 11 h of simulated manipulation and locomotion play, then fine-tuned per task on the 200 synthetic and 20 real episodes. It predicts egocentric frames from the joint states that SONIC and MuJoCo actually reach.
- Rewards. A progress model on DINOv3 features scores each generated video; its increments are the dense reward.
- RL. DICE-RL (a residual on the frozen flow policy with a critic ensemble), ReinFlow and AWR, each run for 15 rounds of 32 fresh rollouts.
Results
- Real robot, 20 trials per task: DICE-RL reaches 85/75/65% on the three tasks, versus 40/20/10% for BC and 60/45/40% for Real RL on the same 20 episodes. RL in the simulator drops success to 30/0/0%.
- Baselines that also use world models score lower: 33.3% mean for VLAW and 40.0% for WMPO.
- Play data improves every world-model metric (+1.17 dB interaction-region PSNR on average). Removing it lowers mean success to 63.3%, sparse rewards to 45.0% and a fixed offline buffer to 58.3%.
Limitations
Each task needs its own world model and policy adaptation, and the G1-specific interfaces rule out zero-shot transfer to other humanoids. One full run takes about 12 h on 8 H100s, three quarters of it video generation.
Videos
Figures



Abstract
Reinforcement learning (RL) can improve vision-language-action (VLA) 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 (WM). 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 WM with deployment observations. The VLA, whole-body controller (WBC), and video WM then form a closed-loop RL environment with dense progress rewards. Across three real robot tasks, play data improves every evaluated WM metric and downstream policy success; video WM RL raises mean success from 23.3% to 75.0%, compared with 48.3% for real robot RL. We therefore leverage the complementary strengths of both environments: simulation provides policy initialization and broad coverage, while the video WM offers a deployment-aligned environment for policy improvement. Project website: https://qianzhong-chen.github.io/wevo.github.io/.
Summaries are written automatically from each paper's text. Figures and videos come from the paper and its project page; all credit goes to the authors.
