Towards Synergistic Human-Robot Co-Adaptation Via Reciprocal Feedback for Shared Contact Tasks

In this work, we propose a human-robot physical interaction scheme designed to facilitate contact-rich manip- ulation tasks. In the proposed framework, neither the robot nor the human agent can complete the task independently, but a shared cost function aligns their efforts and drives them toward success. The robot agent, governed by a reinforcement learning algorithm, can exert forces and modulate its Cartesian impedance while continuously receiving evaluative feedback in the standard RL training paradigm. Simultaneously, the human agent applies forces via a standard PS4 joystick and receives both vibrotactile and visual feedback reflecting task performance. During training, the learning algorithm receives the superposition of its own and the human’s actions, allowing it to implicitly benefit from the human’s rapidly adapting strategy. We hypothesize that human agents can adapt more rapidly than RL and, when provided with feedback grounded in real measurements, can make more quantifiable decisions. During this rapid human adaptation phase, the robot concurrently ac- quires skills from the human, thereby accelerating training and improving overall efficiency. The proposed interaction scheme was evaluated in a realistic simulation environment involving 10 participants. Preliminary results indicate that participants receiving vibrotactile feedback adapted more quickly, enabling the robot to acquire the desired skill in only a few episodes for simple tasks. For more challenging tasks, human-trained RL agents required additional autonomous training, yet still achieved convergence far faster than PPO-only training. This co-adaptive framework combines the complementary strengths of humans and robots, providing a versatile foundation for contact-rich manipulation that may be extended to diverse tasks and robotic platforms.