[robotics-worldwide] [jobs] Social Crowd Nagivation Internships at Honda Research Institute (Peter Trautman)
ibrahimkakbar1 at gmail.com
Wed Mar 11 16:41:36 PDT 2020
Social Crowd Navigation Internship at Honda Research Institute
I have open intern positions at Honda Research Institute (San Jose,
California, USA) in social navigation. My goal for any intern are
publications; relatedly, I strive to enhance the student’s PhD
Currently, We are about to begin a large scale study of social navigation
and trajectory prediction at the University of California at Santa Cruz
with Leila Takayama. We should have a test space up and running by the
summer for the intern to leverage (e.g. onboard people tracking, ground
truth instrumentation of the space for data collection).
Thus, the student would be able to focus on the crowd forecasting and
navigation side without much concern over the perception aspect.
I’m looking for a variety of skill sets—for example, the navigation
(multi-) objective functions I'm currently exploring are non-convex,
meaning an understanding of convex/nonlinear optimization to devise clever
convex relaxations or novel non-convex routines would be encouraged.
Additionally, proposed solutions for generating actions should be real time
(0.1s) to be considered viable and ideally provide some insight about
“closeness” to global optimality.
I’m also open to other formalizations—e.g., game theoretic or interactive
POMDP based navigation techniques.
Recently, (deep) reinforcement learning approaches to social navigation
have seen quite a bit of activity; however, many important questions remain
largely unanswered providing potentially interesting research directions.
For example, what is a “correct” reward function for social navigation?
What are the limitations of a social navigation policy trained in a
simulation (e.g. can training against social forces agents recover
something different than a social forces policy? If so, why? If the RL
agent learns something different than a social forces policy, is this
policy “better” than social forces?)
Similarly, can we recover “social” behaviors using deep learning? If so,
what datasets should we collect? How much data and how many features do we
Those are just a few ideas; I’m open to anything that a student is
passionate about. It's hard to find a human robot interaction topic that
can’t be studied under social navigation.
For further detail the official job description is:
Please apply by sending your CV to ptrautman at honda-ri.com and
peter.trautman at gmail.com.
More information about the robotics-worldwide