Abstract
Humans make extensive use of haptic exploration to map and identify theproperties of the objects that we touch. In robotics, active tactile perceptionhas emerged as an important research domain that complements vision for taskssuch as object classification, shape reconstruction, and manipulation. Thiswork introduces TAP (Task-agnostic Active Perception) -- a novel framework thatleverages reinforcement learning (RL) and transformer-based architectures toaddress the challenges posed by partially observable environments. TAPintegrates Soft Actor-Critic (SAC) and CrossQ algorithms within a unifiedoptimization objective, jointly training a perception module anddecision-making policy. By design, TAP is completely task-agnostic and can, inprinciple, generalize to any active perception problem. We evaluate TAP acrossdiverse tasks, including toy examples and realistic applications involvinghaptic exploration of 3D models from the Tactile MNIST benchmark. Experimentsdemonstrate the efficacy of TAP, achieving high accuracies on the Tactile MNISThaptic digit recognition task and a tactile pose estimation task. Thesefindings underscore the potential of TAP as a versatile and generalizableframework for advancing active tactile perception in robotics.