• Collaboration results with KITECH

This research focuses on the development of an advanced tactile sensing framework for robotic manipulation, aiming to achieve tactile super-resolution through the integration of tactile sensor data, simulation-based learning, and real-world validation. In collaboration with KITECH, the study establishes a closed-loop research pipeline that connects tactile inference, real-to-simulation modeling, deep reinforcement learning, and simulation-to-real transfer.

First, tactile information is acquired from a robotic gripper equipped with a high-density tactile sensing module. The measured tactile signals are processed using learning-based inference models to estimate detailed contact conditions, including pressure distribution, contact location, normal force, shear force, and torque. This enables the robot to obtain richer tactile information beyond the physical resolution of the sensor array.

The inferred tactile data are then used to construct a Real2Sim environment, where real-world grasping interactions are converted into simulation labels and physical parameters. By generating contact-force distributions and interaction labels from experimental data, the simulation environment can more accurately reflect real manipulation conditions.

Based on this tactile-aware simulation environment, deep reinforcement learning is applied to train robotic grasping and manipulation policies. The robot learns how to control its fingers and adjust grasping strategies by using tactile feedback, allowing it to perform stable and adaptive manipulation tasks under various contact conditions.

Finally, the trained policies are transferred back to the real robotic system through a Sim2Real process. The similarity between simulated and real tactile responses is evaluated, and the model is refined to improve the reliability of real-world manipulation. Through this framework, the research aims to enhance robotic perception, improve manipulation stability, and enable intelligent tactile feedback-based control for next-generation robotic hands.