Grasp-Sym: A Neuro-Symbolic Approach to Robust Grasp Classification

Robotic prosthetic hands hold great promise for transradial amputees, assisting them with performing daily life activities. A critical factor for practical robotic prostheses is ensuring that grasp pre-shaping is both accurate and fast. This requires a model capable of making reliable grasp predictions on a variety of objects, including those it has not encountered before. Current literature lacks such analysis and a new definition and evaluation strategy is needed. Moreover, while vision-language models (VLM) have demonstrated strong generalization, their huge parameter space makes them impractical for deployment on embedded devices due to high computational costs and unnecessary information unrelated to grasp estimation. Traditional classification and detection models, on the other hand, offer realtime performance but struggle to generalize to unseen objects. This work presents Grasp-Sym, a method that combines symbolic AI with neural networks to incorporate reasoning ability without relying on large-scale language models. By using expert knowledge to refine grasp predictions, Grasp-Sym enables logical grasp selection based on object shape, size, and affordance part while maintaining efficiency. Experimental results show that Grasp-Sym improves accuracy on unseen objects by 25.1 % compared to SOTA, while maintaining real-time inference speed, making it more suited for practical robotic prostheses.