Exo-Plore: Exploring Exoskeleton Control Space through Human-aligned Simulation

Exoskeletons show great promise for enhancing mobility, but providing appropriate assistance remains challenging due to the complexity of human adaptation to external forces. Current state-of-the-art approaches for optimizing exoskeleton controllers require extensive human experiments in which participants must walk for hours, creating a paradox: those who could benefit most from exoskeleton assistance, such as individuals with mobility impairments, are rarely able to participate in such demanding procedures. We present Exo-plore, a simulation framework that combines neuromechanical simulation with deep reinforcement learning to optimize hip exoskeleton assistance without requiring real human experiments. Exo-plore can (1) generate realistic gait data that captures human adaptation to assistive forces, (2) produce reliable optimization results despite the stochastic nature of human gait, and (3) generalize to pathological gaits, showing strong linear relationships between pathology severity and optimal assistance.

💡 Research Summary

Exo‑plore is a novel simulation framework that combines neuromechanical modeling with deep reinforcement learning (RL) to discover optimal hip‑exoskeleton control parameters without any human‑in‑the‑loop experiments. The authors focus on two tunable parameters of a simplified impedance‑style hip controller: the torque gain κ and the temporal delay Δt, where the assistive torque is τ_exo(t)=κ·u(t‑Δt) and u(t)=sinθ_r−sinθ_l encodes the relative hip motion.

The framework consists of (1) a Gait Data Generator that produces realistic walking trajectories under various assistance conditions, and (2) an Exoskeleton Optimizer that efficiently searches the (κ,Δt) space using a surrogate neural network trained on massive simulation data.

In the generator, a human controller is built from three modules: PoseNet predicts desired joint positions, a PD controller converts these into joint torques, and a Muscle Coordination Network (MCN) maps the torques to muscle activations for a 164‑muscle musculoskeletal model. Training uses a composite reward: gait fidelity (step length, speed, head stability, sway), arm‑movement regularization, metabolic energy minimization, and a novel Human‑Exoskeleton Interaction (HEI) term that encourages the simulated human to reduce resistance to the device, thereby capturing the “resistance‑minimization” hypothesis observed in real subjects.

Metabolic cost is modeled as MEE=∑ m_i α_i a_i^{β_i}, where α and β are tuned so that the simulated walking‑speed versus cost‑of‑transport (CoT) curve reproduces the classic upward‑opening parabola and the preferred walking speed reported in biomechanics literature. An iterative algorithm searches candidate (α,β) pairs, trains the generator without assistance, evaluates the resulting preferred speed, and selects the pair that best matches experimental data; the final values are α=1.5, β=1.0.

After the generator is calibrated, thousands of gait rollouts are performed across a grid of (κ,Δt) values and walking speeds. The resulting CoT measurements train a surrogate network that yields a smooth, differentiable CoT landscape, enabling gradient‑based or Bayesian optimization to locate the minimum CoT efficiently despite the stochastic nature of RL‑generated motions.

Experimental validation shows that Exo‑plore reproduces key human findings: (i) assistive torque and power scale linearly with assistance level and walking speed; (ii) metabolic reductions of roughly 10–15 % are observed, matching published human studies; (iii) the optimal delay Δt decreases as walking speed increases, reflecting faster feedback requirements at higher speeds.

Crucially, the framework generalizes to pathological gaits. By systematically weakening specific muscle groups, the authors create five impairment scenarios. In four of them, the optimal torque gain κ exhibits a strong linear relationship with the severity of the impairment (R² > 0.85), indicating that more severe deficits benefit from higher assistance. The fifth scenario shows a modest deviation, suggesting limits to linearity but still confirming that the simulation captures clinically relevant trends.

Overall, Exo‑plore demonstrates that (1) a neuromechanical‑RL generator can be calibrated to match real human adaptation patterns, (2) a surrogate‑based optimizer can reliably find assistance parameters despite simulation noise, and (3) the approach extends to impaired populations, offering a cost‑effective alternative to exhaustive human experiments for personalized exoskeleton design. This work paves the way for rapid, data‑driven development of assistive devices tailored to individuals who are otherwise unable to participate in lengthy laboratory studies.