Physical interactive robotics, ranging from wearable devices to collaborative humanoid robots, require close coordination between mechanical design and control. However, evaluating interactive dynamics is challenging due to complex human biomechanics and motor responses. Traditional experiments rely on indirect metrics without measuring human internal states, such as muscle forces or joint loads. To address this issue, we develop a scalable simulation-based framework for the quantitative analysis of physical human-robot interaction. At its core is a full-body musculoskeletal model serving as a predictive surrogate for the human dynamical system. Driven by a reinforcement learning controller, it generates adaptive, physiologically grounded motor behaviors. We employ a sequential training pipeline where the pre-trained human motion control policy acts as a consistent evaluator, making large-scale design space exploration computationally tractable. By simulating the coupled human-robot system, the framework provides access to internal biomechanical metrics, offering a systematic way to concurrently co-optimize a robot's structural parameters and control policy. We demonstrate its capability in optimizing human-exoskeleton interactions, showing improved joint alignment and reduced contact forces. This work establishes embodied human simulation as a scalable paradigm for interactive robotics design.
Figures & tables
Fig. 1: Demonstrations of the Digital Human Embodiment with interactive simulation framework. (a) The co-optimization pipeline applied to a wearable exoskeleton, the human policy is trained and applied to robotics optimization. (b) The framework’s scalability to diverse interactive robotics, illustrated by a daily collaborative task with a humanoid robot.
Fig. 2: The human-exoskeleton coupled simulation model. (a) The full system, showing the MS-Human-700 model wearing the OpenExo exoskeleton. (b) A detailed view of the human-robot physical interface, where the adduction joint is passive to allow adaptation to natural leg abduction movement. (c) Illustration of the adjustable structural parameters in the interactive system. Local cuff adjustments correspond to vertical shifts at binding sites, while global assembly module adjustments modify the axis relative to attached human body segment through rotations and translations. The interactions were modeled with compliant elastic tendons elements (gray spheres) that connect the exoskeleton shell to the underlying human segment and transmit assistive forces.
Fig. 3: Flowchart of the co-optimization loop. The optimization algorithm proposes a new set of exoskeleton parameters (control and structure). The human-robot coupled simulation is executed for several gait cycles with these parameters. The resulting motion and physiological data are used to evaluate the cost function. This cost is returned to the optimizer, which updates its internal model and proposes the next set of parameters to evaluate, iterating until convergence.
Fig. 4: Validation of the digital human’s walking kinematics. The simulated joint angles (blue solid lines) are compared with the reference motion capture data (green dashed lines) for major joints. The shaded areas represent the standard deviation across 10 simulation trials with different initialization time steps, indicating high consistency.
Fig. 5: Robustness evaluation of the learned walking policy against external perturbations. (a) The agent’s dynamic recovery sequence after an impulsive force is applied to the pelvis, where the pink arrow visualizes the applied disturbance force. (b) Recovery time across a range of forward (positive) and backward (negative) forces. The red line shows the time required to recover the original kinematic trajectory, demonstrating a fast return to the intended motion. The shaded area represents the standard deviation over 10 randomly initialized trials.
Fig. 6: Performance comparison of optimization strategies. The top three panels illustrate the average joint error, muscle force, and contact force during the co-optimization of controller and structure parameters across episodes, showcasing the improvements. The bottom panel compares the normalized minimum total cost achieved so far for three strategies: controller-only (blue), structure-only (green), and co-optimization (red). The co-optimization approach demonstrates significantly faster convergence and a lower final cost, highlighting the benefits of concurrently optimizing control and structure.
Fig. 7: Enhanced human-robot joint axis alignment after co-optimization. The left column shows the shortest distance between the corresponding human and exoskeleton joint axes, while the right column shows the angle between them. Shaded areas represent the standard deviation across 10 simulation trials. The optimized parameters significantly reduce both distance and angular deviation, enabling more effective and comfortable force transmission.
GenerativeBionics, Genova, Italy. · 2Artificial and Mechanical Intelligence, Istituto Italiano di Tecnologia, Genova, Italy. · School of Computer Science, University of Manchester, Manchester, United Kingdom. +3