Learning in-hand manipulation for a compliant underactuated gripper with interactive human supervision

该项目旨在通过人类演示和交互反馈来设计和开发能够学习自适应顺应性的欠驱动机器人抓取器，以实现对可变形物体的软操作。在第一年，我们完成了第一个目标，即开发一种紧凑型多自由度机器人抓取器，其具有解耦的手腕，能够实现每个关节的独立位置和力控制。我们使用开发的夹持器进行了微创外科应用的实验验证。采集了不同刚性材料抓取时的执行器信号，并训练了无传感器抓持力预测的高斯过程回归模型。我们还设计了一个多目标逆运动学控制框架，可以同时处理多个约束下的关节位置控制。下一个目标是通过强化学习、拟真实技术和交互反馈从人类演示中学习自适应顺应性。我们计划开发一个模拟环境，可以捕获人类演示，并使用最先进的深度强化学习算法训练无模型策略。

This project aims to design and develop underactuated robotic grippers that can learn adaptive compliance for soft manipulation of deformable objects through human demonstrations and interactive feedback.In the first year, we accomplished the first objective of developing a compact multi-dof robotic gripper with a decoupled wrist that enables independent position and force control of each joint. We carried out experimental validation for minimally invasive surgical applications using the developed gripper. We collected actuator signals when grasping materials with different stiffness and trained a Gaussian process regression model for sensorless grip force prediction. We also devised a multi-objective inverse kinematics control framework that can handle joint position control under multiple constraints simultaneously.The next objective is to learn adaptive compliance from human demonstrations using reinforcement learning, sim-to-real techniques and interactive feedback. We plan to develop a simulation environment that can capture human demonstrations and train a model-free policy using state-of-the-art deep reinforcement learning algorithms.