SINDyffuse: Sparse Biomechanical Guidance for Human Motion Generation

Human motion generation models can synthesize diverse actions but often produce physically implausible artifacts, such as foot penetration and asymmetrical poses. Existing physics-based guidance methods can mitigate these issues, but commonly require hand-designed constraints or costly optimization. We introduce SINDyffuse, a diffusion framework trained with a learned, prompt-conditioned biomechanical consistency objective. The objective combines 40 differentiable Rajagopal-model constraint channels with 80 OpenSim MocoTrack-derived muscle-activation targets. A sparse coefficient model selects prompt-relevant relationships between locomotion and contact features and these biomechanical targets, and the resulting loss regularizes the denoiser's predicted clean motion during training. Standard generation retains the classifier-free DDPM sampler and requires neither inverse kinematics nor an OpenSim solve. On HumanML3D, SINDyffuse improves contact-sensitive realism metrics while retaining competitive textual alignment and motion diversity.