From Crowd Motion Prediction to Robot Navigation in Crowds

Poddar, Sriyash; Mavrogiannis, Christoforos; Srinivasa, Siddhartha S. · 2023 · OpenAlex-citations

DOI: 10.1109/iros55552.2023.10341464

archive: archived pipeline: cataloged verified

Get this paper ↗ (DOI — opens at the source; we link to it, we don't host it)

Summary

This paper investigates the extent to which improvements in crowd motion prediction accuracy translate into improved robot navigation performance in crowded environments. While state-of-the-art deep learning models, such as Social GAN (S-GAN), have significantly improved offline pedestrian motion forecasting, their utility for real-world robot navigation remains unclear due to the challenges of transferring performance to close-interaction settings and the limitations of existing simulation benchmarks. The authors specifically ask whether superior prediction models yield safer and more efficient navigation when integrated into a robot’s control system. To address this, the authors integrate S-GAN, a probabilistic trajectory prediction model, into a Model Predictive Control (MPC) framework. They compare this against baselines using simple constant-velocity (CV) and constant-velocity with noise (CVN) predictions. The system was deployed on a self-balancing Honda ballbot in a laboratory setting. The experimental design included three distinct crowd behavior conditions: cooperative (natural walking), aggressive (users ignoring collisions), and distracted (erratic movement). The robot navigated a 3.6 × 4.5 m² workspace, with performance measured by safety (minimum distance to humans) and efficiency (time to goal). The study also included simulated experiments using ORCA-based agents to contrast with real-world results. The results demonstrate that while S-GAN’s prediction accuracy successfully transferred from offline datasets to the real-world robot, this superiority did not translate into better navigation performance. In real-world trials, S-GAN-20 (using 20 samples) achieved lower Average and Final Displacement Errors than CV baselines across all conditions. However, statistical tests showed no significant difference in safety or time-to-goal metrics between the MPC using S-GAN and the MPC using simple CV prediction. In simulations, CV models actually outperformed S-GAN in prediction accuracy because ORCA agents exhibit linear behaviors easily approximated by CV, yet this did not result in superior navigation either. The authors found no statistical correlation between lower prediction error and higher navigation performance. The study concludes that substantial improvements in prediction model accuracy are necessary to yield significant gains in crowd navigation tasks. The current gap between S-GAN and CV predictions, while statistically significant in error metrics, is not large enough to impact the robot’s decision-making in tight spaces. The authors suggest that future work should focus on developing prediction models with even lower errors or incorporating explicit formalisms of model confidence into the control loop to better handle the entanglement of robot and crowd motion.

Provenance

The full processing record for this entry. Every stage of this paper's journey through the pipeline is logged — what ran, with which tool and model, how many attempts it took, and when it last completed.

Summary generated by qwen3.6-27b-prismaquant on 2026-06-26; verification: verified.

Topics

Ranked by relevance to this paper. Hover a topic for its definition.

• anticipation