Proper movements enhance mobility, coordination, and muscle activation, which are crucial for performance, injury prevention, and overall fitness. However, traditional simulation tools rely on strong modeling assumptions, are difficult to set up and computationally expensive. On the other hand, generative AI approaches provide efficient alternatives to motion generation. But they often lack physiological relevance and do not incorporate biomechanical constraints, limiting their practical applications in sports and exercise science. To address these limitations.
We propose a novel framework, BIGE, that combines bio-mechanically meaningful scoring metrics with generative modeling.
BIGE integrates a differentiable surrogate model for muscle activation to reverse optimize the latent space of the generative model
Enables the retrieval of physiologically valid motions through targeted search.
Through extensive experiments on squat exercise data, our framework demonstrates superior performance in generating diverse, physically plausible motions while maintaining high fidelity to clinician-defined objectives compared to existing approaches.
Our framework allows clinicians to impose multiple biomechanical constraints and criteria on the
generated motion, ensuring adherence to physiologically meaningful properties. These constraints
can target various aspects of the motion, including joint kinematics (pelvis tilt, angular velocity ω
), joint center dynamics (center of mass (COM) velocity and acceleration)
To promote physiologically relevant muscle activations, we enforce range constraints on the ‘Vastus Medialis’ muscles. This leads to deeper squats by restricting the activations of the left and right vastus medialis muscles within a user-defined range.
Latent variables (sampled randomly) are decoded using the decoder to generate joint kinematics output. Next, hierarchical transformations are applied to the biomechanical model, to compute joint centers. Then a surrogate model predicts muscle activations. Finally, clinician-defined constraints are imposed on the derived variables.
Comparison of generated samples from baselines and BIGE. The yellow curve represents the movement of the hip joint over the entire squat cycle. BIGE generates a more realistic squat motion compared to baselines.
Comparison of generated samples from MDM and BIGE with the reference data. The yellow curve represents the movement of the hip joint over the entire squat cycle. BIGE generates a more natural squat motion compared to MDM. The red circle highlights the artifact observed in the pelvic tilt for MDM-generated motion.
The generated motion samples are ordered by the peak muscle activation. The red and green lines at 50% squat cycle represent the depth of the squat. Our guidance strategy leads to a more physiologically accurate squat motion as evidenced by the increased depth of the squat.
For the temporal segmentation, that is to find the start and the end frame for each exercise iteration, we found out that the angular velocity of the joints is a good indicator.
The attached video shows an activity being performed. The color of the mesh changes for every iteraction. The top right plot shows the DTW metric similarity between different selected segments. The middle plot in the image shows the angular velocity for each joint ( in deg/s ) for each timestep. The red peaks in the bottom plot showcase the timestep at which an iteration ends.