NASA’s new dark energy space telescope can also detect killer asteroids
technologyreview.com·3h ago
AI Research
arXiv CS.AI·May 25·8 min read
VGAS: Value-Guided Action-Chunk Selection for Few-Shot Vision-Language-Action Adaptation
TL;DR
Vision-Language-Action (VLA) models that combine visual understanding with language reasoning struggle to adapt to new tasks with few examples, often producing actions that look correct but fail due to subtle spatial misalignments. VGAS (Value-Guided Action-chunk Selection) uses a two-stage approach: a fine-tuned VLA generates multiple action candidates, then a specialized Transformer critic called Q-Chunk-Former selects the geometrically most precise option using explicit geometric regularization.
✦ Why It Matters
Engineers building robotic systems can leverage VGAS to reliably adapt vision-language models to new tasks with minimal training data.
Key Takeaways
- Utilize VGAS for improved action selection in VLA models.
- Incorporate Q-Chunk-Former to resolve geometric ambiguities.
- Apply Explicit Geometric Regularization for better action ranking.
- Leverage publicly available code for practical implementation.
How It Works
VGAS operates by first generating action proposals using a fine-tuned VLA model, which ensures a high recall of potential actions. The Q-Chunk-Former then evaluates these proposals, focusing on their geometric accuracy to eliminate near-miss actions that could lead to execution errors.
This dual approach allows VGAS to select action chunks that are both semantically relevant and geometrically sound, enhancing overall task performance.