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

Summary

OmniVLA-RL is a vision-language-action model that adds explicit 3D spatial modeling and online reinforcement learning to robotic manipulation. The paper claims this combination improves action precision, training stability, and benchmark success rates on LIBERO and LIBERO-Plus.

Problem

Existing VLA models often understand language and image semantics well but miss fine 3D spatial details needed for grasping, placement, and obstacle-aware manipulation.
Common spatial fusion designs add 3D features only before or after the core model, which limits joint reasoning over language, vision, spatial cues, and actions.
RL fine-tuning for VLA models is hard to stabilize: PPO is heavy because it needs a value model, and the paper argues GRPO can become unstable when applied at token level.

Approach

The model uses a Mixture-of-Transformers (MoT) backbone with three experts inside shared Transformer layers: a Reasoning Expert for vision-language understanding, a Spatial Expert for 3D scene features, and an Action Expert for control generation.
The Reasoning Expert uses SigLIP to encode multi-view RGB observations and language tokens. The Spatial Expert uses VGGT to extract fine-grained spatial features and adds an auxiliary spatial decoder during training.
A Block-wise Causal Attention mask treats reasoning and spatial tokens as a fully visible prefix, while action tokens form a causal suffix. This lets action generation read scene context without letting action noise corrupt scene understanding.
Actions are generated with Conditional Flow Matching over action chunks, conditioned on spatial, semantic, and language features.
For online RL, the paper introduces Flow-GSPO: it converts the flow-matching action sampler from a deterministic ODE into a stochastic SDE, then applies GSPO at the action-block level so exploration stays stochastic and optimization matches sequence actions better.

Results

On LIBERO, the paper claims an average success rate of 97.6%.
The abstract and introduction say the method surpasses mainstream existing methods on LIBERO and LIBERO-Plus.
The paper claims better convergence speed and final performance than PPO and GRPO baselines on LIBERO-Plus.
The provided excerpt does not include full result tables, per-task scores, or exact PPO/GRPO comparison numbers beyond the 97.6% average success rate on LIBERO.

Link

http://arxiv.org/abs/2604.17706v2