Generating realistic human motion respecting biomechanical constraints and physical plausibility is challenging. ForceDiff is proposed as a physics-guided diffusion framework that models contact forces and biomechanical constraints during the generative process, ensuring physically achievable motions.
Key findings
ForceDiff integrates ground reaction force prediction and physics-based guidance into a unified diffusion architecture.
The model operates in a force-augmented latent space, optimizing joint positions, velocities, and contact forces.
A hierarchical denoising strategy generates coarse motion trajectories, refining them through physics-based projection steps.
Limitations & open questions
The proposed method needs comprehensive validation against state-of-the-art baselines.
Further studies are required to assess the model's scalability and applicability in diverse scenarios.