Reproduce “Trace and Pace: Controllable Pedestrian Animation via Guided Trajectory Diffusion” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Davis Rempe, Zhengyi Luo, Xue Bin Peng, Ye Yuan, Kris Kitani, Karsten Kreis, Sanja Fidler, Or Litany
Researchers at NVIDIA, Stanford University, and collaborators developed a system for generating controllable, realistic pedestrian trajectories and full-body animations in complex environments. It integrates guided diffusion models for high-level planning with physics-based character control, enabling users to specify objectives like waypoints and social interactions while maintaining physically plausible and diverse motions.
Key insights
- Combining a guided diffusion model for high-level trajectory planning with a physics-based reinforcement learning controller for low-level animation enables both naturalness and fine-grained control over virtual pedestrians.
- The proposed "clean guidance" formulation allows for precise, test-time control over diffusion-generated trajectories using arbitrary analytical loss functions, improving adherence to objectives and sample quality.
- Integrating the learned value function from the physics-based controller into the trajectory diffusion model's guidance leads to more "animatable" trajectories that are easier for the character to execute and result in higher fidelity movements.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.