Real-time environmental understanding through active classification using tactile sensing in robotic disassembly

This paper proposes a generalized active classification framework designed to classify the environment of a robot through controlled tactile interaction, with direct application to robotic disassembly of Waste Electrical and Electronic Equipment (WEEE). In this domain, purely vision-based systems frequently fail due to occlusions, dust, and reflective surfaces, making haptic identification an operational necessity. By formulating environment identification as an active sensing problem, we utilize the entropy of learned interaction models and a receding-horizon optimization to select control inputs that maximize information gain from force–torque feedback. Our approach transitions from a passive classifier to an active framework that dynamically probes the environment to resolve belief uncertainty across six distinct tactile classes, demonstrated through the high-precision detection and unfastening of screws using an industrial 6-DOF robotic arm. Experimental results on a consumer microwave demonstrate that the active framework achieves a 51% reduction in detection latency for screws compared to passive baselines and consistently outperforms expert-designed heuristics. Furthermore, the system exhibits robust zero-shot generalization to novel industrial objects, indicating that learned tactile signatures are invariant to the global object context. These results provide a scalable foundation for autonomous disassembly in high-variance environments like electronic waste recycling, while offering a broader methodology for any robotic classification task where control can be leveraged to accelerate information gain.