OmniVLA-RL: A Vision-Language-Action Model with Spatial Understanding and Online RL

TL;DR

OmniVLA-RL is a novel vision-language-action model that improves spatial understanding and reinforcement learning stability through a Mix-of-Transformers architecture and a flow matching reformulation.

Contribution

It introduces OmniVLA-RL with a Mix-of-Transformers design and Flow-GSPO, addressing spatial perception and training stability issues in VLA models.

Findings

Outperforms existing methods on LIBERO benchmarks

Enhances action precision and training robustness

Achieves decent overall performance

Abstract

Visual-Language-Action (VLA) models represent a paradigm shift in embodied AI, yet existing frameworks often struggle with imprecise spatial perception, suboptimal multimodal fusion, and instability in reinforcement learning. To bridge these gaps, we propose OmniVLA-RL, a novel architecture that leverages a Mix-of-Transformers (MoT) design to synergistically integrate reasoning, spatial, and action experts. Furthermore, we introduce Flow-GSPO, which reformulates flow matching as a Stochastic Differential Equation (SDE) process and integrates it with Group Segmented Policy Optimization (GSPO) to enhance action precision and training robustness. Extensive evaluations on the LIBERO and LIBERO-Plus benchmarks demonstrate that OmniVLA-RL achieves decent overall performance and surpasses mainstream existing methods, effectively overcoming the fundamental limitations of current VLA models.