Generative GaitNet

How to read this

Category

Method: deep-RL control of a full-body musculoskeletal model for gait

Contributions

• Generative GaitNet, a deep-RL architecture controlling a full-body musculoskeletal model with 304 Hill-type musculotendons
• A pre-trained, integrated network learned over a 618-dimensional continuous space of anatomy conditions (mass, proportions, bone deformity, muscle deficits) and gait conditions (stride, cadence)
• Generates a range of healthy and pathological gaits in real-time physics-based simulation from anatomy and gait inputs

Context

Builds on muscle-actuated character control, extending the scalable muscle simulation line of Lee et al.'s Scalable Muscle-Actuated Human Simulation and Control toward a generative, condition-parameterized network.Builds on: Scalable Muscle-Actuated Human Simulation and Control

Correctness

Relies on Hill-type muscle modeling and learning across a high-dimensional anatomy-gait space; efficacy and expressiveness are shown through simulated healthy and pathological gaits, so results are demonstrations within the simulated model rather than clinically validated predictions, and coverage at the edges of the 618-D condition space is the caveat.

Clarity

Fairly technical; a first pass conveys the conditioned-generative-control idea, a second pass for the RL setup and the anatomy/gait parameterization.

How to read it

First pass for the scope of the condition space and what generative control over gait means; second pass on the network architecture and RL training if you work on musculoskeletal control or predictive gait simulation.