2D or 3D: Who Governs Salience in VLA Models? -- Tri-Stage Token Pruning Framework with Modality Salience Awareness

TSP-VLA addresses a critical bottleneck in embodied AI: the latency of multi-modal VLA models. With 11 forks and 0 stars within 7 days, this is clearly a fresh research release (likely from a high-output academic lab) where the code is being actively mirrored or tested by peers. The technical moat lies in the 'Modality Salience Awareness' logic, which determines the relative importance of 2D vs 3D data for specific robotic tasks. However, the defensibility is low because token pruning is a standard optimization vector. Frontier labs like Google DeepMind (RT-2/RT-H) or OpenAI/Figure are likely already implementing proprietary versions of cross-modal pruning to maintain high hertz control loops. The project is a valuable contribution to the open-source robotics stack (e.g., as an add-on for OpenVLA), but it faces high displacement risk as VLA architectures shift from standard Transformers to more efficient backbones like Mamba or State Space Models, which handle long sequences (tokens) differently.