Presents DeCAL, a physics-aware tactile fusion architecture that adaptively integrates tactile signals and models contact dynamics to improve dexterous manipulation under occlusion.
The provided search context does not contain information regarding a model or paper titled DeCAL. The available results describe alternative approaches such as TacFiLM (which uses Feature-wise Linear Modulation for lightweight fusion), DreamTacVLA (which employs a tactile world model for future prediction), UniTacVLA (unified understanding and prediction), and TacVLA (contact-aware fusion for robust manipulation).
DeCAL targets a key failure mode of dexterous vision-language-action models: their dependence on visual observation when the hand, object, or surrounding geometry occludes the very contact region that determines task success. The paper introduces a physics-aware tactile fusion architecture that treats tactile sensing as a first-class signal for grounded manipulation rather than as a secondary modality. By adaptively integrating tactile measurements with visual and language-conditioned representations, the system aims to recover contact-relevant state information—such as interaction geometry, force-related cues, and slip or grasp conditions—that is unavailable or ambiguous to cameras alone.
The central technical contribution is a contact-aware latent co-imagination mechanism that couples perceptual fusion with physically informed reasoning about contact dynamics. Rather than simply concatenating tactile and visual inputs, DeCAL appears to model how contact evolves over time and to jointly imagine latent future states involving both visual appearance and interaction consequences. This design is intended to make the action policy more robust in occluded or partially observable settings, where purely visual policies can become brittle because they cannot directly observe the forces, deformations, or contact transitions that govern successful dexterous manipulation.
The work matters because it pushes dexterous VLA models beyond appearance-driven control toward physically grounded manipulation. Tactile sensing has long been important for fine-grained contact tasks, but integrating it into modern multimodal action models in a way that supports prediction and control remains an open challenge. DeCAL’s approach is significant if it demonstrates that tactile-aware latent modeling can improve robustness, contact reliability, and generalization in dexterous manipulation, especially in cluttered or occluded environments where vision alone is insufficient.