Natural Multicontact Walking for Robotic Assistive Devices via Musculoskeletal Models and Hybrid Zero Dynamics

Kejun Li, Maegan Tucker, Rachel Gehlhar, Yisong Yue, Aaron D. Ames

Introduction

Generating stable and natural walking patterns for robotic assistive devices, such as prostheses and exoskeletons, remains a significant challenge in rehabilitation robotics. While mathematical frameworks like Hybrid Zero Dynamics (HZD) ensure stability, they often produce "robotic" gaits that lack biological realism and require extensive manual tuning by experts. This research presents a novel framework that integrates musculoskeletal models directly into the robot's control optimization, bridging the gap between mathematical stability and human biomechanics.

Methods and Modeling

The researchers developed an integrated framework that combines biological muscle models with rigorous control theory to generate anthropomorphic walking gaits.

Musculoskeletal Modeling

The team employed Hill-type muscle-tendon units (MTUs) to model ten major lower-body muscles, including the gluteus, hamstrings, and calves. These models account for the physical limits of muscle force based on length and contraction velocity. By using these biological limits as hard constraints in the optimization process, the system naturally prunes away unnatural-looking solutions that are mathematically stable but biologically impossible.

Musculoskeletal model and joint definitions

Hybrid Zero Dynamics (HZD)

Walking is modeled as a multi-domain hybrid system to account for discrete transitions, such as the shift from double-support to single-support phases. The HZD framework ensures that the resulting gaits are periodically stable and robust.

Integrated Optimization and Learning

The core of the system is an optimization loop that evaluates gait "naturalness" against human motion capture (MoCap) data. To fine-tune high-level parameters like walking speed or step height, the authors utilized Preference-Based Learning (PBL) via the LineCoSpar algorithm. This approach reduces the need for manual "expert-in-the-loop" tuning by systematically finding the best constraints.

Integrated optimization framework

Experimental Validation

The framework was validated on the AMPRO3, a dual-actuated transfemoral prosthesis. Two non-disabled subjects tested the device using a bypass adapter. Performance was measured using motion capture and 6-channel electromyography (EMG) to assess the subjects' actual muscle activity during walking.

Key Results

  • Biomechanical Realism: Gaits generated with musculoskeletal constraints showed a significantly lower Root Mean Square Error (RMSE) relative to human MoCap data compared to standard HZD gaits.
  • User Preference: In blind tests, subjects strongly preferred the musculoskeletal-informed gaits, reporting higher comfort and a more natural walking feel.
  • Reduced Effort: EMG data indicated lower muscle activation in the subjects' own legs when using the "with muscles" gaits. This suggests that the robot's motion was better synchronized with the user's natural intent, reducing the need for compensatory effort.

Conclusion

By embedding the physical constraints of human muscles directly into the robot's control logic, this framework achieves stable, natural walking without the need for exhaustive manual tuning. Future work will focus on extending the system to 3D walking environments and personalizing muscle parameters for individual amputee users to further improve performance and comfort.

Summary updated

BibTeX
@article{li2022natural,
  author = {Kejun Li and Maegan Tucker and Rachel Gehlhar and Yisong Yue and Aaron D. Ames},
  title = {{Natural Multicontact Walking for Robotic Assistive Devices via Musculoskeletal Models and Hybrid Zero Dynamics}},
  journal = {IEEE Robotics Autom. Lett.},
  year = {2022},
  doi = {10.1109/lra.2022.3149568},
}
Figure