TL;DR
This paper introduces a novel data-driven approach to control powered prostheses by learning the human-prosthesis interaction forces from motion capture data. This allows for stable walking control without requiring additional sensors on the user, and the proposed robust controller demonstrates improved tracking performance even with unexpected human movements.
What Problem Does This Paper Address?
Controlling powered prostheses to achieve natural and stable walking is a significant challenge. Traditional methods often rely on extensive tuning or require additional sensors on the user, which can be cumbersome, expensive, and prone to noise. A key difficulty is accounting for the unpredictable interaction forces between the human and the prosthesis, especially during the stance phase where the prosthesis bears the user's weight. Existing model-based controllers struggle to incorporate this human interaction without precise force measurements, leading to limitations in robustness and performance. This paper aims to bridge this gap by developing a model-dependent controller that learns and accounts for human interaction forces directly from motion capture data, eliminating the need for on-board force sensors.
Methods
The core of this work lies in a data-driven methodology that leverages motion capture data to characterize human-prosthesis interaction.
- Motion Capture and Data Processing: Human subjects wearing a powered prosthesis (AMPRO3) are recorded using an Optitrack system. This data captures limb and joint trajectories. This raw data is then processed to extract limb pitch angles, joint angles, and joint positions, projecting the motion onto the sagittal plane.
- Human-Prosthesis Model: A 12-DOF hybrid dynamic model of the human-prosthesis system is developed, incorporating the prosthesis's actuated joints and the human-prosthesis interface. This model is crucial for simulating and understanding the system's dynamics.
- Gait Design and Nominal Trajectory Generation: A gait optimization problem is solved to reconstruct a stable nominal walking gait that closely matches the observed motion capture data. This involves defining state-based outputs and using Bézier curves to parameterize trajectories.
- Interaction Force Estimation and Force Tube Construction: By simulating the human-prosthesis model with the obtained nominal gait and the motion capture trajectories, the interaction forces at the socket are estimated. These estimated forces from multiple walking steps are then used to construct a "force tube"—a bounded range of possible interaction forces—represented by Bézier curves.
- Robust CLF-QP Controller: A robust Control Lyapunov Function (CLF) Quadratic Programming (QP) controller is designed. This controller guarantees convergence to the nominal trajectory while explicitly accounting for the estimated range of interaction forces (the force tube). The controller is formulated to be robust to perturbations in the human's movement and interaction forces, without relying on real-time force sensing.
Datasets
The primary dataset used in this paper is generated from motion capture experiments of a human subject wearing a powered transfemoral prosthesis (AMPRO3). This dataset consists of 37 step cycles of walking data. The authors do not introduce a new benchmark dataset but rather utilize their collected motion capture data to derive a nominal trajectory and interaction force profiles for controller design and simulation.
Key Insights and Ideas
- Learning Interaction Without Sensors: The most significant insight is the demonstration that human-prosthesis interaction forces can be effectively characterized and bounded using only motion capture data and a dynamic model. This bypasses the need for expensive, noisy, and potentially fragile force sensors on the prosthesis itself.
- Force Tube for Robustness: The concept of a "force tube" is a clever way to represent the uncertainty in human interaction. By bounding the possible forces, the controller can be designed to be robust to a wide range of human behaviors, ensuring safety and stability.
- Model Dependence is Key: The paper strongly advocates for model-dependent control in prosthetics. By incorporating knowledge of the human-prosthesis system dynamics and learned interaction forces, the controller achieves superior tracking and robustness compared to model-independent or less model-dependent approaches.
- Bridging Simulation and Reality: The methodology aims to create a simulation environment that accurately reflects real-world human-prosthesis interaction, enabling the development and testing of robust controllers before experimental deployment.
Key Results
- Stable Nominal Trajectory: The gait optimization procedure successfully generated a stable nominal walking trajectory that closely matched the motion capture data.
- Force Tube Characterization: The method successfully estimated interaction force profiles and constructed a force tube that bounded these forces across multiple walking steps.
- Robust Controller Performance: Simulation results showed that the proposed robust CLF-QP controller achieved stable tracking of the nominal trajectory even when subjected to perturbed human trajectories and interaction forces.
- Superiority over Other Controllers: The robust CLF-QP controller outperformed both a robust-passive controller and a PD controller in tracking performance under perturbed conditions. Notably, the CLF-QP controller did not require extensive gain tuning, unlike the robust-passive controller.
- Smoothed Control Input: Modifications to the QP formulation (using torque bounds and CLF relaxation) successfully smoothed the prosthesis control input, making it more physically feasible for implementation.
Limitations and Future Work
- Simulation-Based Validation: The primary limitation is that the controller's performance is validated through simulations, not experimental deployment on a physical prosthesis. The authors acknowledge that real-world experiments are needed to fully assess tracking performance, energy efficiency, and robustness.
- Impact Model Simplification: The paper notes that the rigid impact model used in simulations can lead to discrete jumps in velocities, which are not as pronounced in real human walking due to inherent compliance.
- Single Subject Data: The current methodology is based on data from a single subject. Future work aims to test the controller across multiple subjects and investigate the generalizability of force profiles, potentially predicting forces without motion capture for new users.
- Experimental Realization: The authors plan to experimentally realize the controller on a physical prosthesis platform.
- Force Sensor Integration: If a force sensor were to be incorporated, the collected force profile statistics could be used for designing better force sensor filters.
Summary updated