A rehabilitation robot control framework with adaptation of training tasks and robotic assistance

Jiajun Xu, Kaizhen Huang, Tianyi Zhang, Kai Cao, Aihong Ji, Linsen Xu, Youfu LiFrontiers in Bioengineering and Biotechnology 2023 · DOI

TL;DR

This paper presents a novel rehabilitation robot control framework that simultaneously adapts both the training trajectory and the level of robotic assistance based on a patient's real-time performance. By using electromyography (EMG) signals to estimate muscle torque, the system ensures that patients remain actively engaged—neither overwhelmed by difficulty nor passive due to excessive robotic help.

What Problem Does This Paper Address?

In robot-assisted rehabilitation, the "challenge point" is critical: if a task is too hard, a patient with severe motor impairment may become frustrated or unable to complete it; if it is too easy, or if the robot provides too much help, the patient becomes passive, which hinders neurological recovery.

Current state-of-the-art systems often struggle with two things:

  1. Static Trajectories: Most robots follow a pre-set path. While some allow for slight deviations via physical human-robot interaction (pHRI), they don't intuitively "reshape" the future path to accommodate a patient's specific movement intentions.
  2. Fixed Assistance: Many robots don't effectively "step back" when a patient shows improvement. To trigger neuroplasticity, the robot should provide only the minimum assistance necessary—a concept known as "Assist-as-Needed" (AAN).

Methods

The authors developed a dual-layer control architecture to solve these issues:

  • EMG-Driven Torque Estimation: The system captures raw EMG signals from key leg muscles (like the gluteus maximus and rectus femoris). These signals are fed into a Hill-type musculoskeletal model to estimate the user's intended joint torque. This allows the robot to "know" how much effort the human is actually contributing.
  • Trajectory Deformation Algorithm: Instead of a rigid path, the robot uses a high-level controller that deforms the desired trajectory in real-time based on pHRI. It utilizes a minimum-jerk model to ensure that even when the path changes, the motion remains smooth and compliant, preventing jerky movements that could injure a patient.
  • Adaptive AAN Controller: The low-level controller regulates robotic assistance by modifying feedback gains ($K_D$). If the patient’s performance (estimated via EMG) improves, the controller reduces the gain, allowing for more tracking error and forcing the patient to work harder. Conversely, if the patient struggles, the gain increases to provide more support.

Datasets

The study did not use large-scale public datasets. Instead, the framework was validated through:

  • Biomechanical Simulation: The AnyBody Modeling System (AMS) was used to provide "ground truth" torque data for optimizing the EMG-driven muscle model.
  • Experimental Validation: Data was collected from three healthy subjects using a custom-built lower extremity rehabilitation robot. While the subjects were healthy, they were used to simulate varying levels of participation to test the controller's adaptability.

Key Insights and Ideas

  • Biological Feedback Loop: The core "aha moment" is the integration of biological signals (EMG) directly into the task-planning phase. The robot doesn't just react to force; it reacts to the intent measured at the muscle level.
  • The Balance of Error: The paper argues that "perfect" tracking is actually bad for rehab. By intentionally allowing for a bounded tracking error, the robot creates "space" for the patient’s voluntary effort.
  • Smoothness as Safety: By using a trajectory deformation approach rather than simple admittance control, the system maintains high-order continuity (smoothness), which is vital for patient comfort and safety during gait training.

Key Results

The proposed framework was compared against a standard admittance control scheme. The quantitative improvements were significant:

  • Movement Smoothness: The Dimensionless Squared Jerk (DSJ) was orders of magnitude lower for the proposed system ($7.66 \times 10^8$) compared to admittance control ($3.78 \times 10^{12}$), indicating much smoother transitions.
  • Active Participation: Normalized EMG values were nearly double in the proposed framework ($0.87$ vs $0.48$), proving that the AAN controller successfully encouraged subjects to exert more effort.
  • Tracking Accuracy: Despite being "adaptive," the system maintained a lower Root Mean Square Error (RMSE) of $0.07$ rad, compared to $0.19$ rad for the baseline.
  • Clinical Potential: Predicted Fugl-Meyer Assessment (FMA) scores—a standard clinical metric—were significantly higher for the proposed framework ($30.4$ vs $18.8$).

Limitations and Future Work

  • Small, Healthy Sample: The primary limitation is that the system was only tested on three healthy subjects. The dynamics of neurologically impaired gait (e.g., spasticity or tremors) may present challenges not seen in this pilot.
  • Task Variety: The current study focused on standard gait cycles. Future work needs to explore more diverse and complex training tasks.
  • AI Integration: The authors suggest that machine learning could be used in the future to better estimate individual motor capabilities and further personalize the assistance levels.

Summary updated

Abstract

Robot-assisted rehabilitation has exhibited great potential to enhance the motor function of physically and neurologically impaired patients. State-of-the-art control strategies usually allow the rehabilitation robot to track the training task trajectory along with the impaired limb, and the robotic motion can be regulated through physical human-robot interaction for comfortable support and appropriate assistance level. However, it is hardly possible, especially for patients with severe motor disabilities, to continuously exert force to guide the robot to complete the prescribed training task. Conversely, reduced task difficulty cannot facilitate stimulating patients' potential movement capabilities. Moreover, challenging more difficult tasks with minimal robotic assistance is usually ignored when subjects show improved performance. In this paper, a control framework is proposed to simultaneously adjust both the training task and robotic assistance according to the subjects' performance, which can be estimated from the users' electromyography signals. Concretely, a trajectory deformation algorithm is developed to generate smooth and compliant task motion while responding to pHRI. An assist-as-needed (ANN) controller along with a feedback gain modification algorithm is designed to promote patients' active participation according to individual performance variance on completing the training task. The proposed control framework is validated using a lower extremity rehabilitation robot through experiments. The experimental results demonstrate that the control scheme can optimize the robotic assistance to complete the subject-adaptation training task with high efficiency.
BibTeX
@inproceedings{xu2023rehabilitation,
  author = {Jiajun Xu and Kaizhen Huang and Tianyi Zhang and Kai Cao and Aihong Ji and Linsen Xu and Youfu Li},
  title = {{A rehabilitation robot control framework with adaptation of training tasks and robotic assistance}},
  booktitle = {Frontiers in Bioengineering and Biotechnology},
  year = {2023},
  doi = {10.3389/fbioe.2023.1244550},
}
Figure