Adaptive Control of an Upper Extremity Rehabilitation Robot with Backlash

Borna Ghannadi, John McPheeProceedings of the International Conference of Control, Dynamic Systems, and Robotics 2016 · DOI

Abstract

In Canada, over 40,000 new stroke cases are reported annually, costing the healthcare system approximately $3.2 billion per year. Robotic rehabilitation has emerged as a vital tool for delivering the repetitive practice necessary for motor recovery. However, safety concerns in human-robot interaction often restrict control scenarios to admittance or impedance strategies, which require highly accurate dynamic models. Traditional robust or adaptive controllers often struggle with mechanical discontinuities like gear backlash, requiring either high gains or "persistently exciting" inputs that are unsuitable for recovering patients. This paper presents a model reference adaptive control (MRAC) approach designed to estimate and compensate for backlash in real-time, ensuring smooth and safe therapeutic interaction without requiring jerky movements from the user.

Introduction

Stroke rehabilitation relies on the brain's neuroplasticity, which is best stimulated through high-intensity, repetitive physical therapy. While robots are ideally suited for this task, the mechanical reality of robotic hardware introduces significant control challenges. One of the most pervasive issues is backlash—the clearance or "play" between mating components such as gears and belts.

In a rehabilitation context, backlash creates discontinuities that degrade the transparency of the robot. If the controller cannot accurately account for this play, the patient may experience a "clunky" or unnatural resistance, which can impede the recovery process and decrease patient motivation. This research focuses on developing a control architecture that compensates for these mechanical imperfections while prioritizing the safety and comfort of the human user.

The Challenge of Backlash and Adaptive Control

Standard control strategies for rehabilitation robots, such as impedance control, rely on a precise dynamic model to make the robot feel "invisible" or provide specific levels of assistance. When backlash is present, the relationship between the motor's position and the robot's actual end-effector position becomes non-linear and uncertain.

Adaptive controllers are often used to "learn" these uncertainties on the fly. However, a major theoretical requirement for most adaptive systems is a persistently exciting (PE) input. To accurately identify mechanical parameters, the system usually needs to be driven through a wide range of rapid, high-frequency movements. For a stroke patient with limited mobility, performing these "exciting" movements is often impossible and clinically undesirable. High-gain robust controllers offer an alternative but can lead to instability or vibration, posing safety risks during human-robot interaction.

Proposed Methodology

The researchers at the University of Waterloo developed a Model Reference Adaptive Controller (MRAC) specifically designed to handle backlash without the need for PE inputs. The approach is divided into three primary phases:

1. Initial Parameter Identification

Before the patient interacts with the robot, the system undergoes an initial identification phase. Using a modified homotopy parameter identification method, the researchers establish a baseline model of the robot’s dynamics in a controlled environment where human interference is absent.

2. Adaptive Backlash Compensation

Once a user begins therapy, the dynamics change. Human interaction effectively alters how backlash manifests within the system's kinematics. The controller employs an adaptation law—derived from the foundational work of Tao and Kokotovic—to update the estimates of left and right gear clearances in real-time.

3. Kinematic Correction vs. Parameter Convergence

The core innovation of this approach is the decision to prioritize kinematic correction over total parameter convergence. By focusing on compensating for the immediate physical effects of backlash rather than trying to perfectly identify every underlying mechanical constant, the system remains stable and effective during the smooth, slow movements typical of stroke therapy. This eliminates the need for the jerky, "persistently exciting" inputs that hinder traditional adaptive systems.

Experimental Validation

The proposed controller was tested using a specialized upper extremity rehabilitation robot developed through a collaboration between Quanser Inc. and the Toronto Rehabilitation Institute (TRI).

To validate the theory, the researchers utilized a simplified single-link model that mimicked the joint drive backlash found in more complex systems. This controlled environment allowed for the precise measurement of how well the adaptation law could handle discontinuous dynamics. The results demonstrated that the modified MRAC could successfully update backlash parameters in real-time, providing a smoother interaction for the user even when the mechanical "slop" was significant.

Discussion and Future Work

While the preliminary results on the single-link model are promising, several steps remain before this technology reaches clinical settings. The current study highlights a critical trade-off: by sacrificing the mathematical ideal of parameter convergence, the researchers gained a system that is significantly more viable for medical use.

Future research directions include: * Scaling to Multi-DOF Systems: The next phase involves applying the adaptive controller to the full 2-degree-of-freedom (DOF) belt-driven parallelogram robot. * Clinical Trials: Evaluating the "feel" of the robot with actual patients to determine if the backlash compensation leads to better engagement and improved motor recovery outcomes. * Addressing Parameter Drift: Long-term stability must be monitored to ensure that the lack of PE inputs does not lead to "drift" in the estimated parameters over extended therapy sessions.

Conclusion

By integrating a modified adaptation law with a focus on user comfort, this research provides a path forward for more sophisticated and user-friendly rehabilitation robots. Compensating for mechanical backlash without requiring erratic movements ensures that the technology remains a help, rather than a hindrance, to the patient's journey toward recovery.

Summary updated

Abstract

In Canada over 40,000 new stroke cases are reported annually [1], and it costs the Canadian health care system $3.2 billion a year Thus, new rehabilitation technologies are being investigated. A rehabilitation robot can be used to deliver repetitive practice to the stroke patients, which is the key element for motor recovery Safety issues regarding humanrobot interactions restrict selection of control scenarios. Currently admittance and impedance control approaches and their variations are used to control the rehabilitation robots To implement these control strategies, a complete and accurate dynamic model of the robotic system is required. This issue can be addressed by incorporating robust or adaptive control approaches in the above strategies. For the robust control, if the dynamic uncertainties of the robot are too great, the quality of adaptive assistance or resistance may be compromised during therapy In the adaptive control, for the convergence of the adaption law, a persistently exciting input is required In addition, both adaptive and robust controllers may need high gains when mechanical discontinuities (such as gear backlash) are introduced to the system dynamics. Hence, a precise dynamic model estimation is essential, and can be done through a proper system parameter identification approach.
BibTeX
@inproceedings{ghannadi2016adaptive,
  author = {Borna Ghannadi and John McPhee},
  title = {{Adaptive Control of an Upper Extremity Rehabilitation Robot with Backlash}},
  booktitle = {Proceedings of the International Conference of Control, Dynamic Systems, and Robotics},
  year = {2016},
  doi = {10.11159/cdsr16.120},
}
Figure