KINESIS: Motion Imitation for Human Musculoskeletal Locomotion
A new reinforcement learning framework learns from 1.8 hours of human motion data to generate physiologically plausible muscle activity.
A team of researchers from EPFL and Harvard has introduced KINESIS, a novel reinforcement learning (RL) framework designed to tackle the complex challenge of human musculoskeletal locomotion. Unlike previous torque-controlled humanoid models, KINESIS directly addresses key biomechanical constraints, including joint limits and the non-linear, overactuated nature of musculotendon control. The model-free system was trained on just 1.8 hours of human locomotion data, yet it achieves strong performance on unseen motion trajectories. Through a technique called negative mining, KINESIS learns robust locomotion priors that enable its deployment on diverse downstream tasks.
KINESIS demonstrates remarkable scalability, seamlessly controlling biomechanical models of increasing complexity, from simpler skeletons to systems with up to 290 individual muscles. Crucially, the AI learns to generate muscle activation patterns that show a strong correlation with real human electromyography (EMG) data, a significant step toward physiological plausibility. This capability allows the framework to be applied to practical applications like converting text commands into motion, navigating to target points, and executing dynamic skills such as football penalty kicks. The research, accepted for presentation at ICRA 2026, provides code, videos, and benchmarks, positioning KINESIS as a powerful new tool for advancing studies in human motor control, rehabilitation robotics, and more lifelike animated characters.
- Controls highly complex musculoskeletal models with up to 290 individual muscles, far surpassing simpler joint-angle controllers.
- Generates muscle activity that correlates with human EMG data, a key benchmark for biological accuracy in simulation.
- Trained on only 1.8 hours of motion data and applied to tasks like text-to-control and executing football penalty kicks.
Why It Matters
Enables more realistic humanoid robots, advanced prosthetics, and better digital humans for film, games, and medical research.