Figure 2.

Gait-kinematics for RL agents for all models during walking and running

Shown are the hip, knee, ankle and GRF values averaged over 5 rollouts of 10 s walking and running on flat ground. We excluded rollouts that did not achieve the whole episode length to clearly highlight the achieved kinematics. For walking, we observe slight discrepancies between experimental data (gray) and the RL behaviors (red), which are bigger for high-dimensional models.

(A) The experimental data show human subjects walking at $1.2$^m/_s and is included in SCONE.⁴⁴

(B) The experimental data show human subjects running at $5$^m/_s and was extracted from Hamner et al.⁵⁷ For running, the behaviors deviate overall more from experimental data, a weak knee flexion can be observed. Nevertheless, the characteristic single-peaked GRF curve is present for all models, except for the MyoLegV0. The proposed reward function provides a strong and flexible starting point for researchers aiming to create robust and natural controllers for high-dimensional musculoskeletal systems. Also, see the videos on the website. In all simulations, the values (red) are averaged over both legs and 5 rollouts of 10 s are recorded. The experimental data (shaded) is shown as maximum and minimum values over all subjects.