Session: 07-10-03: Medical Robotics, Rehabilitation, and Surgery III

Paper Number: 167097

Imitation Learning-Based Adaptive Control for a Quasi-Serial Rehabilitation Robot Using Behavior Cloning 

This study presents a novel imitation learning-based control framework using Behavior Cloning (BC) for a quasi-serial robotic system designed to deliver adaptive assistance during upper-limb rehabilitation. The robot, combining both serial and parallel kinematic structures, is well-suited for therapeutic applications due to its high back-drivability, safety, and mechanical compliance. Traditional controllers like PID offer simplicity and acceptable tracking but fail to adapt to dynamic and patient-specific motor behaviors. On the other hand, more complex model-based controllers like MPC demand accurate system modeling and computational resources, which limits their real-world clinical applicability.

To address these shortcomings, the authors develop a data-driven BC approach. BC learns control policies by imitating expert demonstrations, providing context-aware and patient-tailored assistance without explicit modeling of the patient-robot interaction. The BC policy is trained on synthetic expert trajectories generated by combining multiple PID controllers (with varying gain settings) and a simplified MPC-inspired strategy within a simulated expert human-in-the-loop environment. This hybrid "doctor-in-the-loop" design mimics therapist behavior—ranging from conservative to aggressive assistance—and forms a diverse, expert-like dataset.

The rehabilitation robot is modeled as a five-link quasi-serial structure with three active degrees of freedom (DOFs). Its control architecture is implemented in a high-fidelity Python simulation, supporting both linear and half-circular trajectory tracking in 3D space. To ensure robustness, the system includes modules that introduce realistic disturbances such as abrupt patient-applied torques and stochastic noise.

The BC training dataset includes state-torque pairs from six expert controller configurations (three PID variants integrated with MPC-like behavior), under both disturbance and noise scenarios. Each expert policy operates based on Cartesian error inputs, mapping to joint torques through Jacobian transpose methods. The resulting BC policy is trained via supervised learning using a neural network with two hidden layers, mapping a 16-dimensional state vector (including current and past errors) to a 3-dimensional torque output. The BC controller is evaluated against traditional PID control across multiple conditions: Linear trajectory with abrupt torque, Linear trajectory with sensor noise, Half-circle trajectory with abrupt torque, Half-circle trajectory with sensor noise.

The results indicate that BC consistently outperforms PID across all tasks and metrics. For trajectory tracking, BC achieves lower Root Mean Square Error (RMSE) in all scenarios. For instance, on the half-circle trajectory with abrupt disturbances, BC records an RMSE of 0.059 compared to PID’s 0.143.

Furthermore, BC excels in generating smoother motion. Analysis of jerk (the derivative of acceleration) reveals that PID introduces frequent and large spikes, especially during motion transitions. In contrast, BC maintains a consistently low jerk magnitude, contributing to smoother, more physiologically natural movement—an essential aspect of effective rehabilitation.

Torque analysis shows that BC outputs smoother and more stable torque profiles than PID, avoiding abrupt spikes and oscillations. Additionally, the BC controller demonstrates more consistent “assist-on” behavior—sustaining support when needed without rapid on-off switching as seen in PID.

The torque-error relationship also supports BC’s efficiency: it generally requires less torque to correct comparable errors, suggesting better utilization of assistive effort. This aligns with the assist-as-needed rehabilitation philosophy, which aims to encourage patient effort while only providing support when necessary.

The discussion section emphasizes BC’s benefits in terms of robustness, smoothness, and adaptive assistance. While PID is a reliable baseline, its rigid response lacks the flexibility required for patient-specific rehabilitation. BC, on the other hand, learns expert-like strategies that generalize well under perturbed conditions. The synthetic expert demonstrations, although not collected from actual therapists, reflect a rich variety of clinical intentions by combining responsive PID behavior with predictive MPC traits.

From a clinical standpoint, smoother movements and intelligent assistance timing are crucial for patient comfort, engagement, and neuroplastic recovery. BC’s ability to produce low-jerk trajectories and sustain support during challenging motion phases could lead to improved rehabilitation outcomes.

Nonetheless, the authors acknowledge limitations. The work is currently restricted to simulations and a simplified patient model. Real-world testing on physical hardware and with diverse patient profiles is necessary to validate clinical efficacy. Moreover, the BC policy is task-specific; generalization to other tasks may require fine-tuning or retraining.

Future work may involve integrating online learning or reinforcement learning to adapt policies in real-time, expanding trajectory types, refining assist logic, and exploring transfer learning across patients or conditions. Additionally, introducing a continuous torque-scaling assist mode based on patient effort could further personalize therapy.

Presenting Author: Soroush Korivand Mississippi State University

Presenting Author Biography: Dr. Soroush Korivand is an Assistant Professor in the Michael W. Hall School of Mechanical Engineering at Mississippi State University. His research focuses on human-robot interaction, reinforcement learning, and AI-powered robotics, with applications in assistive technologies, rehabilitation, and industrial automation. Dr. Korivand’s work integrates machine learning, computer vision, and neurophysiological signal processing to develop intelligent robotic systems that enhance human capabilities in both healthcare and manufacturing settings.

Authors:

Soroush Korivand Mississippi State University
Seyed Hooman Hosseini-Zahraei Ferdowsi University of Mashhad