Bayesian and Neural Inference on LSTM-Based Object Recognition From Tactile and Kinesthetic Information

Recent advances in the field of intelligent robotic manipulation pursue providing hands with touch sensitivity. Haptic perception encompasses sensing modalities encountered sense (e.g., tactile and kinesthetic sensations). This letter focuses on multimodal object recognition proposes analytical data-driven methodologies to fuse tactile- kinesthetic-based classification results. The procedure is as follows: a three-finger actuated gripper an integrated high-resolution sensor performs squeeze-and-release Exploratory Procedures (EPs). images information acquired using angular sensors finger joints constitute time-series datasets interest. Each temporal dataset fed Long Short-term Memory (LSTM) Neural Network, which trained classify in-hand objects. LSTMs provide estimation posterior probability each given corresponding measurements, after fusion allows estimate through Bayesian inference approaches. An experiment 36-classes carried out evaluate compare performance fused, tactile, systems.The results show that Bayesian-based classifiers improves capabilities for outperforms Neural-based approach.