Safe Diffusion Model Predictive Control for Interactive Robotic Crowd Navigation

Sepehr Samavi, Anthony Lem, Fumiaki Sato, Sirui Chen, Qiao Gu, Keijiro Yano, Angela P. Schoellig, Florian Shkurti

Keywords: social navigation, diffusion, model predictive control

Abstract:

To navigate a crowd safely without collisions, robots need to inter-
act with humans by predicting potential future motion and duly reacting. While
learning-based prediction models have proven to be a promising approach towards
generating human trajectory forecasts, utilizing these models in a robot controller
presents the challenges of accounting for the coupling of planned robot motion
with human predictions and ensuring that both predictions and robot actions are
safe. Towards addressing these challenges, we present a receding horizon crowd
navigation method for single-robot multi-human environments. We propose a dif-
fusion model to generate multi-modal human trajectory forecasts that we use to
parameterize a Bilevel Model Predictive Control (MPC) problem, which jointly
filters the predictions to satisfy safety constraints (lower-level) and solves for a
robot plan (upper-level) that is coupled with the filtered predictions. We evalu-
ate the open-loop trajectory prediction performance of our diffusion model on the
commonly used ETH/UCY benchmark, and analyze the closed-loop performance
of our robot navigation method in extensive real-robot experiments to demonstrate
safe and efficient robot motion.

Submission Number: 26