Controlling pedestrian crowd dynamics has attracted increasing attention due to its potential impact to save lives in emergency. The "fast-is-slower" effect defines the phenomenon of jamming at an exit or a bottleneck, caused by people rushing to the exit. Originated from the transportation community, it is known that modification of pedestrian facilities can increase efficiency and safety. For example, adding "obstacles" can stabilize flow patterns and make the flow more fluid; adding zigzag-shaped geometries and columns can reduce pressure in panicked crowds. However, modification of infrastructure is often expensive and not easily reconfigurable in real time. In this project, we have proposed a new robot-assisted pedestrian flow regulation scheme that uses a mobile robot to dynamically interact with moving pedestrians for the purpose of achieving flow regulation in real time.

We first conducted robot experiments to characterize human-robot interaction (HRI). We deployed a mobile robot moving in a direction perpendicular to the pedestrian flow in a uni-directional exit corridor, and installed a pedestrian motion tracking system to record the collective motion. We analyzed both individual and collective motion of pedestrians, and measured the effect of the robot motion on the overall pedestrian flow. The experimental results have shown the effect of passive HRI, where the pedestrians' overall speed was slowed down in the presence of the robot, and the faster the robot moves, the lower the average pedestrian velocity becomes. Experiment results have shown qualitative consistency of the collective HRI effect with simulation results using existing social force models.

Based on the passive HRI, we formulated the robot-assisted pedestrian regulation problem as to find the optimal robot motion frequency such that the average pedestrian speed reaches a desired level. We proposed an adaptive dynamic programming (ADP) based learning algorithm that takes the measurement of pedestrian flows from a surveillance camera as input, and outputs the motion frequency for the robot. The ADP method solves the optimal control online and is the type of feedback control. The ADP learning architecture consists of a critic network and an actor network. Utilizing function approximators of neural networks, the actor network generates the control signal, and the critic network estimates the value function that is the sum of discounted rewards from the current time to the infinite horizon future.  The error function of the critic network is defined as the temporal-difference (TD) error. When the TD error is zero, the Bellman Optimality Equation is solved.

We have studied a complex environment based on a real-world scenario that occurred in Mina/Makkah 2006, where hundreds of people were killed in a crowd stampede incident when pedestrian flows from two perpendicular directions merge together through a bottleneck. To regulate the merging pedestrian flows and achieve optimal flow performance, we have developed an ADP based robot-assisted pedestrian regulation algorithm that utilizes passive HRI and tunes robot motion parameters in real time. Special attention was paid to monitor the crowd pressure, the quantity measuring the critical crowd conditions that may evolve into crowd accidents. Simulation results in robotic simulators have demonstrated that our approach regulates pedestrian flows to an optimized outflow by online learning from the real-time observation of the pedestrian flow, and the critical crowd pressure is reduced to prevent potential crowd disasters.

To address the challenge of feature representation of pedestrian motion using the ADP based method, we have further developed an end-to-end robot motion planner for the robot-assisted pedestrian flow regulation problem. It consists of a deep neural network that models the mapping from the image input of pedestrian environments to the output of robot motion decisions. The robot motion planner is trained end-to-end using a deep reinforcement learning algorithm, which avoids hand-crafted feature detection and extraction, thus improving the learning capability for complex dynamic problems. Simulation results have demonstrated that the robot is able to find optimal motion decisions that maximize the pedestrian outflow in different flow conditions, and the pedestrian-accumulated outflow increases significantly compared to cases without robot regulation and with random robot motion.

The project integrates research projects with education activities through robot centric undergraduate and graduate education. The research results of the project have enriched a few fundamental courses on cross-disciplinary subjects with new cutting edge techniques in assistive robotics technology. The project have also provided direct research training to several graduate students and summer undergraduate students who have been trained in robotics research from algorithm design, simulation testing, and field experiments. The research results have been published in journals and international conference proceedings. The PIs have made conference presentations and invited talks on the research, and disseminated the research results to the robotics and control communities.

Last Modified: 11/29/2019
Modified by: Yi Guo