SICNav-Diffusion: Safe and Interactive Crowd Navigation with Diffusion Trajectory Predictions
Sepehr Samavi, Anthony Jia-Hao Lem, Fumiaki Sato, Sirui Chen, Qiao Gu, Keijiro Yano, Angela P. Schoellig, Florian Shkurti
AI summary
Problem
Robots struggle to navigate crowds safely because they must account for interactive coupling between their actions and human predictions, handle multimodal futures, and guarantee collision avoidance without sacrificing efficiency.
Approach
The method uses a joint diffusion model to forecast multi-agent trajectories, which are then filtered in real-time by a bilevel MPC planner that optimizes the robot path while enforcing collision-avoidance constraints via ORCA.
Key results
- JMID outperforms state-of-the-art joint prediction models on the ETH/UCY benchmark
- Bilevel SICNav planner successfully couples multimodal predictions with hard safety constraints in simulation
- 240 real-robot experiments demonstrate safe, efficient, and reactive navigation
- Successful deployment across 8 km of outdoor real-world operations
Why it matters
Provides a reliable framework for deploying autonomous robots in dense human environments where safety and interaction are critical.
Abstract
To navigate crowds without collisions, robots must interact with humans by forecasting their future motion and reacting accordingly. While learning-based prediction models have shown success in generating likely human trajectory predic- tions, integrating these stochastic models into a robot controller presents several challenges. The controller needs to account for interactive coupling between planned robot motion and human predictions while ensuring both predictions and robot actions are safe (i.e. collision-free). To address these challenges, we present a receding horizon crowd navigation method for single- robot multi-human environments. We first propose a diffusion model to generate joint trajectory predictions for all humans in the scene. We then incorporate these multi-modal predictions into a SICNav Bilevel MPC problem that simultaneously solves for a robot plan (upper-level) and acts as a safety filter to refine the predictions for non-collision (lower-level). Combining planning and prediction refinement into one bilevel problem ensures that the robot plan and human predictions are coupled. We validate the open-loop trajectory prediction performance of our diffusion model on the commonly used ETH/UCY benchmark and evaluate the closed-loop performance of our robot naviga- tion method in simulation and extensive real-robot experiments demonstrating safe, efficient, and reactive robot motion. Code: github.com/sepsamavi/safe-interactive-crowdnav.git