MoTVLA: Stimulating the Spatial Reasoning Ability of VLA Models via Multimodal Token Embedding

MoTVLA is a Vision-Language-Action (VLA) model based on the Mamba architecture. It solves the problem of traditional VLA models lacking an explicit spatial verification mechanism through Gaussian Spatial Tokenizer (GST) and Depth-Aware Chain-of-Thought (DA-CoT). It achieves an average success rate of 90% on the LIBERO benchmark while maintaining real-time inference speed on a single GPU.

Traditional VLA models encode visual observations into flat 2D image patch tokens, lacking inherent geometric structure information. Adding monocular depth only provides distance information and cannot express key spatial attributes such as surface orientation and geometric confidence, leading to the lack of an explicit spatial verification mechanism in the policy network and limited performance in high-precision manipulation tasks.

1. Gaussian Spatial Tokenizer (GST)：Converts frozen affine-invariant depth estimation and semantic image patch features into 3D Gaussian primitives (including metric residual mean, diagonal log covariance, and learned opacity), and focuses on geometrically significant regions via spatial attention pooling；2. Depth-Aware Chain-of-Thought (DA-CoT)：Generates four types of structured spatial thinking: 3D object localization, grasp affordance contact geometry, pairwise metric distance, and coarse SE(3) waypoints；3. Mamba-SSM Inference Core：Fuses GST tokens, language tokens, and CLIP features；4. Flow Matching Action Expert：Decodes 16-time-step 7-degree-of-freedom action blocks via dual cross-attention.

• Explicit geometric representation: 3D Gaussian primitives (anisotropic) are more suitable for complex geometric scenes than implicit feature learning；- Spatial chain-of-thought: Extends CoT to spatial reasoning, improving interpretability；- Performance balance: 90% success rate on LIBERO benchmark + real-time inference on a single GPU；- Ablation experiments: GST and DA-CoT contribute independently to performance, and their combination produces a superadditive effect.

• Precision manipulation tasks: Assembly, grasp planning, tool use, collaborative manipulation；- Interpretable robot learning: Analyze reasoning chains, identify spatial understanding blind spots；- New paradigm for multimodal learning: Fusion of continuous geometric information (Gaussian fields) and discrete symbolic reasoning (chain-of-thought), providing references for fields such as autonomous driving and augmented reality.

Limitations：Relies on frozen depth estimation (errors affect spatial representation), computational overhead needs optimization, task generalization remains to be tested；Future Directions：End-to-end Gaussian learning, dynamic scene expansion, cross-robot transfer, human-robot collaboration.

MoTVLA addresses the spatial reasoning limitations of traditional VLA models through GST and DA-CoT, balancing accuracy, efficiency, and interpretability. Its open-source implementation provides a reference for the research community. As robot learning moves toward practical applications, such methods will play an important role.