What Goes Bump in the Night: Learning Tactile Control for Vision-Occluded Crowd Navigation

Kyle Morgenstein, Junfeng Jiao, Luis Sentis

Keywords: Social Navigation, Contact-based Social Navigation, Motion Planning, Reinforcement Learning

TL;DR: We utilize contact force sensing to learn a local planner for safe, vision-occluded social navigation in dense crowds.

Abstract: Broadening the scope of social environments in which robots can be reliably deployed requires understanding how to safely navigate in contact-prone environments. While collision-free social navigation is well studied, navigation planners that incorporate safe contacts remain largely unexplored. Traditional social navigation schemes require the robot to stop short if a collision is imminent. “Freezing” the robot while navigating in a crowd may cause people to trip and fall over the robot, causing more harm than the collision itself. In very dense social environments where contacts are common (e.g. public transit, narrow corridors, doorways, etc.), this control scheme would render the robot stationary and unable to traverse the environment, which would in turn prevent robots from ever being deployed in densely populated locales. Thus, if we wish to deploy robots in crowded human spaces, planning safe contacts is imperative. A further challenge is that people in the most crowded environments will occlude traditional exteroceptive sensors that are closer than the sensor’s minimum resolving distance. To overcome these limitations, we propose a learning-based motion planner and control scheme to navigate ultra-dense social environments using safe contacts for an omnidirectional mobile robot. We formulate the local planner as the solution to a multi-task reinforcement learning problem. We consider the task as 1) follow a coarse set of waypoints from a global planner, and 2) minimize discomfort to humans. We define discomfort as the ratio between measured contact forces and the average pain threshold for blunt impacts between a human and a robot, as established by the ISO 15066:2016 standard. The policy is evaluated in simulation over 160 trials with crowd densities varying between 0 and 1.75 people per square meter. We achieve a 100% safety factor in crowds up to 1.0 people per meter squared, and 90% in crowds with densities greater than 1.0 people per meter squared. Our navigation scheme is able to use contact to safely navigate in simulated crowds of higher density than has been previously reported.

Confirmation: I am a student and I would like to be considered for the funding award.