Reproduce “Unleashing More Actions via Action Compositional Training for VLA Models” on OpenResearch

About this paper

Kai Peng, Jie Lu, Xiaojiang Peng

The ACT-VLA framework, developed by researchers at Shenzhen Technology University, significantly improves the compositional generalization of Vision-Language-Action (VLA) models by automatically synthesizing novel training demonstrations. This approach yielded absolute gains of up to 52.7 and 49.0 percentage points on out-of-distribution compositional tasks within the LIBERO benchmark, enabling VLA models to learn complex skill recombinations without additional human teleoperation.

Key insights

• VLA models implicitly encode compositional structures in their latent representations that can be effectively exploited for automated data generation, not just inference-time steering.
• Internalizing compositional transitions into the model's weights through training-time data synthesis provides a more robust and effective strategy for OOD generalization than relying on per-step latent manipulation during inference.
• Expanding the effective training distribution with high-quality, synthetically generated compositional data without additional human effort significantly enhances VLA generalization and mitigates overfitting.

Summary by alphaXiv.