Beyond Motion Imitation: Is Human Motion Data Alone Sufficient to Explain Gait Control and Biomechanics?

Abstract

With the growing interest in motion imitation learning (IL) for human biomechanics and wearable robotics, this study investigates how additional foot-ground interaction measures, used as reward terms, affect human gait kinematics and kinetics estimation within a reinforcement learning-based IL framework. Results indicate that accurate reproduction of forward kinematics alone does not ensure biomechanically plausible joint kinetics. Adding foot-ground contacts and contact forces to the IL reward terms enables the prediction of joint moments in forward walking simulation, which are significantly closer to those computed by inverse dynamics. This finding highlights a fundamental limitation of motion-only IL approaches, which may prioritize kinematics matching over physical consistency. Incorporating kinetic constraints, particularly ground reaction force and center of pressure information, significantly enhances the realism of internal and external kinetics. These findings suggest that, when imitation learning is applied to human-related research domains such as biomechanics and wearable robot co-design, kinetics-based reward shaping is necessary to achieve physically consistent gait representations.

Figures (7)

• Figure 1: Comparison of (a) a biomechanics inverse dynamics pipeline and (b) a motion-only imitation learning (IL) pipeline integrating a reinforcement learning agent with forward dynamics; ground reaction force measurements are omitted in MOIL approaches.
• Figure 2: Models used for the inverse dynamics and imitation learning. The Visual3D inverse dynamics model and the MuJoCo forward dynamics model shared identical anthropometric configurations (segment lengths, masses, and inertial properties) to ensure consistency.
• Figure 3: Schematic of the actor-critic architecture for PPO-based residual-force control.
• Figure 4: Reward ablation scheme and corresponding reward curves. (a) Schematic of the training protocol, where agents are pre-trained for 700 episodes with kinematic reward only ($R^{\mathrm{k}}$) before entering a 200-episode fine-tuning phase under four conditions: (i) $R^{\mathrm{k}}$ only, (ii) $R^{\mathrm{k}}$ with GRF reward ($R^{\mathrm{grf}}$), (iii) $R^{\mathrm{k}}$ with CoP reward ($R^{\mathrm{cop}}$), and (iv) all terms combined. (b) Average reward values plotted individually for each reward component during the fine-tuning interval (700 to 900 episodes), corresponding to the active terms in each condition.
• Figure 5: Joint angle trajectories and tracking accuracy across ablation conditions. (a) Mean hip, knee, and ankle joint angles over the gait cycle (GC) for the forward-simulated kinematics generated from different ablation conditions. GT indicates ground truth kinematics trajectories. Flex and ext denote flexion and extension, respectively; D-flex and P-flex denote dorsiflexion and plantarflexion of the ankle joint, respectively. (b) The mean and standard deviation of joint angle RMSE for each joint and condition. A dashed horizontal line marks the reference error level reported by Chiu et al. chiu2025speedadaptive. Error bars denote standard deviations.