World2Act: Latent Action Post-Training via Skill-Compositional World Models
TLDR
World2Act transfers world model dynamics to VLA policies via latent-space alignment, improving success rates on simulation and real robots without pixel supervision.
Reasoning
The paper introduces a novel latent-space post-training method that avoids pixel-space supervision, showing clear gains on multiple benchmarks and a real robot. Strengths include the contrastive alignment approach and empirical validation; weaknesses are limited scope of benchmarks and lack of detailed analysis of failure cases.
Read-first score
Read-first score 49.8, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 46.
Field roles
Rank sensitivity
Stability: volatile; rank range: 320.
Keyword Scores
Deep Analysis
Innovations
- Latent-space post-training framework for VLA policies that avoids pixel-space supervision
- Shared video-action latent space induced by contrastively aligning WM-dynamics latents with action embeddings
- Guiding policy action representations toward WM-imagined dynamics instead of decoded pixels
Methodology
World2Act operates in two stages: first, it induces a shared video-action latent space by contrastively aligning WM-dynamics latents with action embeddings; second, it post-trains the VLA by guiding policy action representations toward WM-imagined dynamics rather than decoded pixels. The framework is built on GR00T-N1.6 and evaluated on simulation benchmarks (RoboCasa, LIBERO, Bridge-SIMPLER) and a real robot.
Key Results
World2Act achieves absolute success-rate gains of up to +2.5% on simulation benchmarks and +6.7% on a real robot over finetuned VLA baselines. It outperforms pixel-space WM supervision by up to +6.0%, including on LIBERO where pixel supervision degrades the baseline.