Paper 2601.14133
TwinBrainVLA: Unleashing the Potential of Generalist VLMs for Embodied Tasks via Asymmetric Mixture-of-Transformers
- Published
- Jan 2026
- Research lab
- Independent
- Citations
- 12
- GitHub
- 29 stars
01 In brief
Summary
TwinBrainVLA is a novel Vision-Language-Action (VLA) framework that addresses catastrophic forgetting in VLM fine-tuning for robotic control.
It uses two isomorphic VLM pathways: a frozen 'Left Brain' (generalist) preserving pre-trained semantic knowledge, and a trainable 'Right Brain' (specialist) for motor control.
An Asymmetric Mixture-of-Transformers (AsyMoT) mechanism lets the Right Brain dynamically query and fuse semantic features from the Left Brain, conditioning a flow-matching action expert.
Experiments on SimplerEnv, RoboCasa, LIBERO, and real-robot tasks show significant gains over baselines, with Qwen3-VL-4B achieving 64.5% on SimplerEnv and 54.6% on RoboCasa, surpassing Isaac-GR00T-N1.6 by +7.4% and +7.0% respectively.
Ablations confirm the importance of freezing the Left Brain and dense interaction.
A 'Twin-to-One' distillation shows the dual-stream approach can improve single-stream models.
02 From the paper
Abstract
The fundamental premise of Vision-Language-Action (VLA) models is to harness the extensive general capabilities of pre-trained Vision-Language Models (VLMs) for generalized embodied intelligence. However, standard robotic fine-tuning inevitably disrupts the pre-trained feature space, leading to "catastrophic forgetting" that compromises the general visual understanding we aim to leverage. To effectively utilize the uncorrupted general capabilities of VLMs for robotic tasks, we propose TwinBrainVLA, which coordinates two isomorphic VLM pathways: a frozen generalist (also called "Left Brain") and a trainable specialist (also called "Right Brain"). Our architecture utilizes a Asymmetric Mixture-of-Transformers (AsyMoT) mechanism, enabling the Right Brain to dynamically query and fuse intact semantic knowledge from the Left Brain with proprioceptive states. This fused representation conditions a flow-matching action expert for precise continuous control. Empirical results on SimplerEnv and RoboCasa benchmarks demonstrate that by explicitly retaining general capabilities, TwinBrainVLA achieves substantial performance gains over baseline models in complex manipulation tasks.