$ω$-EVA: Envision, Verify, and Act with Latent Interactive World Models
TLDR
ω-EVA introduces a latent interactive world model that enables an Envision-Verify-Act loop for embodied action generation without generating future videos.
Reasoning
The paper presents a novel three-stage framework that uses action-conditioned latent dynamics and a tri-branch refiner to improve policy proposals. Its strengths include a compact model (1.2B parameters) and strong simulation results, but it lacks real-world experiments and does not address video generation or model-based RL explicitly.
Read-first score
Read-first score 52.2, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 41.
Field roles
Rank sensitivity
Stability: volatile; rank range: 361.
Keyword Scores
Deep Analysis
Innovations
- Envision-Verify-Act loop for embodied action generation using a latent interactive world model
- Tri-branch refiner that jointly reasons over current state, proposal-conditioned future, and proposed action
- Latent space consequence reasoning that avoids generating future videos at inference
Methodology
ω-EVA employs a three-stage framework: first, it learns action-conditioned latent dynamics; second, it trains a language-conditioned flow policy on dynamics-aware visual representations; third, it feeds the policy's proposal back through the world model. A tri-branch refiner then jointly reasons over the current state, proposal-conditioned future, and proposed action to produce the final action chunk, all within latent feature space.
Key Results
Evaluations across diverse single-arm, bimanual, long-horizon, and perturbed simulation settings show that the complete interaction pipeline consistently improves the proposal policy, and latent diagnostics indicate meaningful action-conditioned future structure. The model achieves competitive performance with approximately 1.2B parameters and no additional robot-data pretraining.