Helping Robots Gently Navigate Through Thick Crowds

On December 5, 2017, citing safety concerns for pedestrians, the City of San Francisco decided  to severely restrict the comings and goings of all those delivery robots that have recently started plying the sidewalks. The restrictions include a confinement to industrial areas, a speed limit of 3 mph, a maximum of nine robots in the entire city and the requirement for a human monitor nearby.
The episode illustrates the mounting concerns over the cohabitation between robots and humans in public spaces. A rather timely initiative against this backdrop, Crowdbot gathers 5 universities and 2 industrial partners for a 42-month research project meant to help robots peregrinate smoothly through thick crowds.
“The topic raises lots of questions indeed, scientist Julien Pettré remarks. Is it relevant to have robots amid a crowd? Is it desirable? Is it dangerous? Can the risk be minimized? As far as we are concerned, we believe that, yes, robots' presence may be relevant in some circumstances, the best example being the autonomous wheelchair. For, if it is banned from certain public spaces, then the impaired people using this vehicle are de facto excluded from these locations. On what ground should they be prohibited from patronizing a concert venue or whatnot? Now, as for the pizza delivery bot, it's another debate ” which the lawmaker ultimately will have to arbitrate. One of the aims of the project is precisely to help governments formulate a legislation on the whole issue. “Even though the ethical and legal ramifications have not yet been resolved, one must delineate the technical and scientific questions in order to ascertain what are the risks and conversely the guarantees that can be provided. ”

Robot Freeze

“Today, when a human gets within 10 cm of a robot, the machine is programmed to stop at once in order to avoid collision. That's what we call robot freeze, which, in and of itself, can be counterproductive at times. If the robot travels in a densely-populated area, it will remain at a standstill most of the time, thus becoming ineffective. To make matters worse, if an emergency evacuation occurs, this idle robot could even turn into a obstacle in narrow passages or near the exits toward which people are rushing. ”
All things considered, it might therefore be a better idea to keep robots moving. “We hold that when a robot is meant to act and move in a place shared with humans, such as a public building, its navigation system should be designed in such a way that the robot is able to continue its task or adapt itself to the evolving situation. ” This premise however bumps against a pesky stumbling block: “we have to take into account the fact that the density of people walking around public spaces is not controlled. In other words, the crowd might turn out to be really thick. And even if it is sparse in an area, people might well all flock to one particular spot. ” So no doubt there will be times when autonomous robots will have to meander through dense throngs.

Sensing

To meet the challenge, CrowdBot will unfurl research efforts in several directions. One of them is sensing. “The robot is equipped with tracking cameras as it needs an estimate of the surrounding crowd. There are two elements to this: individual tracking of people who are closest to the robot, as well as an estimate of the collective motion within a defined ‘visibility’ distance. The expertise in this field of computer vision is provided by the RWTH Aachen University in Germany . ”
Another research axis regards navigation. “It's a three-tier strategy. Long term: I perceive a dense crowd in front of me. So I will try to circumvent these people by anticipating the most suitable trajectory to avoid contact. Medium term: I've tried my best but there are people now passing by me. After observing how they move, I will perform local adaptations accordingly. Short term: a bump is about to occur or is even happening. I must react right away. ”
Specialized in robot navigation in dynamic environments, ETH Zurich will work on the long- and medium-term aspects. In Lausanne, EPFL will concentrate on the short term strategy in collaboration with Locomotec, a German company whose cuyBot robot will serve as a testbed. “EPFL has a strong expertise in physical interactions between robots and humans, in particular in the context where these latter two are asked to perform manipulation tasks in a collaborative fashion. The project will provide an opportunity to transfer this knowledge to the realm of mobile robotics. As for Locomotec, they are interested in testing these ideas. Their robot will be equipped with force sensors, thus enabling us to study the forces exerted during robot-crowd interaction. ”

Man-Robot Interaction

Although minimizing collision risks remains the overall goal, researchers are interested in further exploring to what degree gentle bumps could be part of the equation. “If we can ascertain what are the tolerable forces during a man-robot interaction and if we have the measurements and the algorithms that enable the robot to react accordingly, then the machine could seek not to go beyond what is acceptable and behave as to avoid hurting anyone. ”
Two other platforms will be available for the experiments: “one is Pepper, a human-shaped robot on wheels designed by SoftBank Robotics Europe and meant to move in close proximity with humans. The other one is the semi-autonomous wheelchair developed by University College London . Here at Inria, my colleague Marie Babelalso works on this latter field and the project should strengthen her ties with UCL . ”

Simulation

An essential ingredient cementing these technological bricks is crowd simulation, an expertise mostly provided by Inria. Working from the data collected by the robot, “simulation will help predict the evolution of the crowd state in view of anticipating collisions. People are here. They're moving in this direction. They should be there in 10 seconds. So I'll go this way. Maybe could I even manage to predict what will be the people's reaction to my own move.”
Simulation will also enable to assess various scenarios and the risks they entail as preliminary validation stage before evaluating them in real conditions. “It will tell us for instance that one given navigation algorithm provokes say 40 collisions whereas another one generates 20, or results in harsher collisions. ” In this regard, scientists need a new generation of tools which they intend to deliver. “Today's crowd simulators have no anchor in reality. We will have to  extend these techniques to enable them to incorporate a simulated robot, as well as a physical simulation layer to report on collision forces. This will be a complete novelty in the field. ”