AI

HaWMPO Down-Weights Hallucinated World-Model Rollouts

Robb Harlan 4 min read

World-model RL is cheap until the dream lies. HaWMPO, from Jilin University and JD’s Joy Future Academy, is a September 9 paper that tries to catch those lies before they update the policy.

The method post-trains a VLA inside an action-conditioned world model, then runs a hallucination-aware model (HAM) over each generated 8-frame chunk. A Reward-Soft term scales the chunk reward by (1 − α · H). High-H chunks still exist. They just count less.

On LIBERO Spatial / Goal / Object, HaWMPO’s average success is 63.7%, 15.0 points over one-shot SFT OpenVLA-OFT (48.7%) and 2.8 over the authors’ reproduced WoVR* (60.9%). Spatial is the big lift: 77.2% versus 69.2%.

Unitree G1 arms on a white desk with a mesh organizer, tissue box, and objects for a pick-and-place task
Real G1 setup for Tissue-to-Box. Source: HaWMPO paper, Figure 3.

Score the dream, then GRPO

The world model is an action-conditioned Wan2.2-TI2V-5B. It sees five RGB frames (one fixed reference plus four recents) and an 8-step action chunk, then predicts eight future frames at 256×256. The VLA rolls the next chunk on those frames. Eight parallel imagined environments feed GRPO. Training ran on 8 NVIDIA H200s.

HAM fuses the predicted chunk, the actions, the current frame, and the rollout’s first frame through a frozen DINOv3 encoder and a four-layer fusion transformer. Supervision is a composite of DINOv3 similarity, depth consistency, optical-flow trajectory consistency, and MUSIQ. Against 50 human-labeled LIBERO chunks (18 hallucinated, 32 not), HAM’s AUROC is 0.9375. MUSIQ alone is useless.

The authors are honest about the proxy: task failure is not hallucination, and the four metrics are not physics.

HaWMPO pipeline: world-model rollouts, hallucination scores, Reward-Soft, and GRPO updates
Hallucination-aware VLA RL pipeline. Source: HaWMPO paper, Figure 1.

Two G1 chores

Hardware is a G1. Two tasks, 20 trials each, pedal off at eval:

MethodTissue-to-BoxHeadphone-on-StandAverage
SFT OpenVLA-OFT75% (15/20)60% (12/20)67.5%
WoVR80% (16/20)65% (13/20)72.5%
HaWMPO85% (17/20)75% (15/20)80.0%
G1 real-world rollouts: tissue into a box on top, headphones onto a stand on the bottom
Physical G1 executions, not world-model dreams. Source: HaWMPO paper, Figure 9.

That is a 12.5-point bump over the base policy. The paper calls it preliminary. It is two tasks.

Penalty strength α is task-dependent. Object and Spatial like 0.3. Goal likes 1.0. Strip HAM (and Reward-Soft with it) and Spatial drops 8.0–11.6 points depending on step count.

A Human’s Take

I like that they treated world-model RL as a lying intern, not an oracle. Down-weighting the pretty-but-wrong rollouts is the grown-up move. 80% on tissue-in-box is still a lab chore, and Goal did not love the same α as Spatial. If you cannot trust the dream equally across task families, you do not have a generalist trainer. You have a knob.

Sources