Dissertation Title: Predictive Prosthesis Control via Inferred Foot Placement: A Unified Design for Dynamic Locomotion

Abstract:

This thesis introduces an inference-based predictive control framework for noninvasive, accessible prosthesis locomotion assistance. The control framework uses wearable inertial measurement units with a transformer-based model trained through behavior cloning to continuously infer prospective foot-placement targets and forecast kinematics, from which projected time to landing is derived. These signals provide a high-level description of the intended locomotor plan without requiring implanted interfaces or direct neural measurements. The predictive controller uses this information, together with a three-link planar walking model (3LP) and identified Sunfish prosthesis dynamics, to anticipate the intended next-step location and deliver dynamic torque assistance. This formulation is designed to enable movement support that is difficult to obtain from conventional finite-state-machine impedance-control systems, whose observability and control structure limit assistance to predefined gait states. Across evaluation data collected under deployment conditions, prospective foot-placement prediction error was characterized at 0.17–0.25 m. Additionally, anticipatory commands became available at the start of gait steps a median of 0.81 s before cadence-based estimates.

To translate this framework into a deployable system, this work develops a holistic hardware–controller co-design methodology in which mechanical dynamics and control objectives are designed together rather than purely sequentially. Using the same gait dataset as input, this unified approach identifies mechanical and control solutions that improve energy efficiency relative to traditional sequential design strategies. It also enables the discovery of control strategies that leverage system dynamics across time and allows offline trajectory-optimization solutions to be reproduced through online receding-horizon model-predictive control. To facilitate energy storage with clean dynamics, a compact parallel-spring mechanism incorporates a cam-based nonlinear-cancellation method formulated as an optimal-control problem over the mechanism’s configuration; the resulting cam path preserves linear behavior to within 5% across its operating range. System identification using a voice-coil actuator, factor-graph maximum-a-posteriori estimation, and nonparametric stochastic methods provides the modeling, cross-validation, and verification foundation for these designs, enabling the automatic generation and evaluation of optimal controllers, state estimators, and a virtual load cell for external-torque estimation without an additional physical load cell. The virtual load cell agreed with an independent load cell to within approximately 5% and produced physiologically plausible stance-phase torque estimates during untethered walking. These methods are realized in Sunfish, a powered prosthesis platform with unique architectural features for improved human factors and clean dynamics, and Moonfish, an electronics, software, and control foundation demonstrated across neural and traditional prosthesis-control studies.

To evaluate the total system, a preliminary human study provided an end-to-end demonstration on an obstacle course for a single participant with a unilateral transtibial amputation. Relative to the commercial Empower prosthesis, the proposed control system running on the Sunfish prosthesis produced more symmetric ankle kinematics across most course elements. Using ankle range-of-motion asymmetry, defined as the absolute log-ratio between prosthetic- and intact-side values, median asymmetry was 0.108 with the proposed system compared with 0.170 with the Empower across the course, and 0.129 compared with 0.421 during the precision foot-placement task, where lower values indicate greater symmetry. Together, these contributions provide an initial demonstration of an integrated approach to prospective foot-placement inference, predictive control, and prosthesis hardware for dynamic locomotion assistance, providing a noninvasive foundation for future predictive and data-driven control, neural control, and hybrid approaches that combine them.

Hugh Herr
Professor of Media Arts and Sciences
MIT Media Lab

Alexander Slocum
Professor
MIT Mechanical Engineering

Maani Ghaffari
Associate Professor
University of Michigan