TwinBrainVLA: Unleashing the Potential of Generalist VLMs for Embodied Tasks via Asymmetric Mixture-of-Transformers
This work addresses the challenge that fine-tuning general-purpose vision-language models (VLMs) for robotic control often leads to degradation of semantic understanding and interference with fine-grained motor learning. To mitigate this, the authors propose TwinBrainVLA, an architecture that freezes a pre-trained VLM as the “left brain” to preserve open-world semantic comprehension, while introducing a trainable, embodied perception-specific model as the “right brain” to learn high-precision continuous actions. The two components are integrated via an Asymmetric Mixture-of-Transformers (AsyMoT) mechanism and further enhanced by a Flow-Matching action expert module, enabling effective fusion of high-level semantics and low-level control. Experiments demonstrate that the approach outperforms existing methods on SimplerEnv and RoboCasa benchmarks, significantly alleviates catastrophic forgetting, and maintains the pretrained VLM’s general visual understanding capabilities.