Robot autonomous navigation has been improving for decades, allowing robots to get around safely in a large variety of environments, often using only onboard sensing and computation. These improvements are leading to commercial applications such as last-mile delivery, patrol, retail, and logistics. However, as researchers start to consider real-world navigation scenarios another problem has become the center of attention: how can robots navigate in environments populated by humans? When sharing space with people, robots have to modify their behavior: following the shortest path to a goal without collision is not the only objective, but doing it in a way that “respects” the activities the other (human) agents are performing. The rules for respecting other human agents are subtle and often unwritten, involving concerns as perceived privacy, personal safety, predictability, and pedestrian conventions. Moving with respect for others is the core of social navigation. Acknowledging this increasing importance, we would like to start a dialogue with the “social navigation community” to help define better and promote the field; this document and our planned event are our first steps towards it.

We are interested in addressing a range of questions regarding how we define, measure, and implement social navigation. In particular, we seek answers to the following questions:

1. How do we define social navigation?

   • What is a social agent, what properties are we talking about? Safety, politeness, predictability, …?

   • Is a social agent expected to behave as a human, or have an asymmetric behavior, e.g. a servant who accommodates people at all times?

   • Does a social agent need to communicate, e.g. use head movement, lights, or even voice?

   • To what extent does a human need to collaborate with the agent? In other words, is socialness defined without considering human behavior? What if the person behaves erratically?

2. How do we measure social navigation? What are good metrics?

   • Computed metrics? What can they capture? Can they be robustly computed in sim and real?

   • Role of human raters? First-person rating vs third person rating?

3. How do we set up benchmarks for social navigation? How to best set up a repeatable, scalable, and informative benchmark in simulation and real?

   • How to set up a real benchmark? How controlled should it be? Is it reasonable to expect it to be repeatable? Can one open-source it, have one benchmark shared across the community, or replicated in different labs?

   • What is the role of simulation? What aspects of socialness can we test?

The goal of this event is to bring a diverse set of researchers with an interest in Social Navigation and have a closed-form discussion on the above topics. The hope is that the community can arrive at recommendations and a process to benchmark Social Navigation. A potential outcome of this meeting is a technical report with our findings.

Video Recordings from the Event: Introduction and Participants' Presentations

Video Recordings from the Event: discussion roundtables

Participating Researchers and Institutions

Schedule

8:00 - 8:10 - Introduction by the organizers [video]

Alexander Toshev, Robotics@Google

Roberto Martin-Martin, Stanford U (incoming UT Austin)

8:10 - 10:10 - Presentations by each participant group [video]

• Stanford team (Roberto Martin-Martin, Claudia D'Arpino, Eric Li)

• Google team (Alexander Toshev, Soeren Pirk, Anthony Francis)

• MIT team (Jonathan How, Michael Everett)

• Yale team (Marynel Vazques, Nathan Tsoi)

• Honda RI (Pete Trautman)

• Everyday Robots + George Mason (Xuesu Xiao)

• UT Austin team (Peter Stone, Justin Hart)

• University of Maryland (Dinesh Manocha)

• EPFL (Alexandre Alahi)

• CMU (Henny Admoni)

• LAAS-CNRS team (Rachid Alami, Phani Teja Singamaneni, Anthony Favier)

• Georgia Tech team (Sehoon Ha, Noaki Yokohama)

• Aston University (Luis J Manso)

10:10 - 10:20 - Break: brew coffee/tee :)

10:20 - 10:50 - How do you define Social Navigation? [video]
Moderator: Claudia Perez D'Arpino

10:50 - 11:20 - How do you measure Social Navigation? [video]

Moderator: Chengshu (Eric) Li

11:20 - 11:50 - How do you operationalize a Social Navigation Benchmark? [video]

Moderators: Fei Xia and Anthony Francis

11:50 - 12:00 - Summary and Debriefing of the discussions

Alex Toshev and Roberto Martin-Martin

Participant Guidelines

• Introduction by each participant / group: prepare a presentation of the research of your team: what is your research about and how does it connect to social navigation. Keep it short, 8 minutes per team!

• Discussions for the three questions: each group brings 1-2 slides for each of the questions (how to define, measure and benchmark Social Navigation?). We suggest using this template (please, create a copy instead of editing it). We will try to keep it VERY BRIEF: the idea is to bring all options to the table and then open for discussions on the alternatives. No need for elaborated slides, just your ideas!

Further discussions over google group: given the brief time available for our symposium, we may continue unfinished discussions over our google group. Consider joining! Be careful! We have observed that some of the notifications from the group are filtered as spam by Google, check your spam folder.

Organizers