Sara Pohland and Claire Tomlin

EECS Department, University of California, Berkeley

Technical Report No. UCB/EECS-2022-36

May 4, 2022

http://www2.eecs.berkeley.edu/Pubs/TechRpts/2022/EECS-2022-36.pdf

An important challenge in human-robot interaction is the design of socially compliant robot navigation policies that enable safe navigation around crowds of people. Deep reinforcement learning is a popular and effective method to predict human motion and plan paths that avoid collisions with humans while following social norms. While various deep reinforcement learning methods have proven effective for generating crowd-aware navigation policies, these policies generally assume the robot is operating around humans in a large open environment, which is not reflective of typical indoor spaces. To design a socially compliant robot navigation policy that works effectively in complex indoor spaces with walls and other stationary objects, I combined a deep reinforcement learning policy with a global path planning algorithm and a custom safety controller. Combining all these elements in a modular framework, I enabled a robot to reach its goal from an arbitrary starting position, while limiting close encounters with humans and avoiding collisions with humans and stationary objects. I found that my policy achieves an overall success rate of over 99\% when tested in a diverse set of simulation environments comprising different geometries and distributions of humans and stationary obstacles. When compared against a baseline navigation policy that does not utilize learning, I found that my modular approach results in better navigation performance and greater compliance to human social norms. I also implemented my approach on a physical robot to navigate real-world indoor spaces with various humans and stationary objects. These simulation and real-world results demonstrate the value of using learning-based methods as a single component of a larger framework to develop navigation policies that are effective in real-world settings and comfortable for people.

Advisors: Claire Tomlin

---

BibTeX citation:

@mastersthesis{Pohland:EECS-2022-36,
    Author= {Pohland, Sara and Tomlin, Claire},
    Title= {A Modular Framework for Socially Compliant Robot Navigation in Complex Indoor Environments},
    School= {EECS Department, University of California, Berkeley},
    Year= {2022},
    Month= {May},
    Url= {http://www2.eecs.berkeley.edu/Pubs/TechRpts/2022/EECS-2022-36.html},
    Number= {UCB/EECS-2022-36},
    Abstract= {An important challenge in human-robot interaction is the design of socially compliant robot navigation policies that enable safe navigation around crowds of people. Deep reinforcement learning is a popular and effective method to predict human motion and plan paths that avoid collisions with humans while following social norms. While various deep reinforcement learning methods have proven effective for generating crowd-aware navigation policies, these policies generally assume the robot is operating around humans in a large open environment, which is not reflective of typical indoor spaces. To design a socially compliant robot navigation policy that works effectively in complex indoor spaces with walls and other stationary objects, I combined a deep reinforcement learning policy with a global path planning algorithm and a custom safety controller. Combining all these elements in a modular framework, I enabled a robot to reach its goal from an arbitrary starting position, while limiting close encounters with humans and avoiding collisions with humans and stationary objects. I found that my policy achieves an overall success rate of over 99\% when tested in a diverse set of simulation environments comprising different geometries and distributions of humans and stationary obstacles. When compared against a baseline navigation policy that does not utilize learning, I found that my modular approach results in better navigation performance and greater compliance to human social norms. I also implemented my approach on a physical robot to navigate real-world indoor spaces with various humans and stationary objects. These simulation and real-world results demonstrate the value of using learning-based methods as a single component of a larger framework to develop navigation policies that are effective in real-world settings and comfortable for people.},
}

EndNote citation:

%0 Thesis
%A Pohland, Sara 
%A Tomlin, Claire 
%T A Modular Framework for Socially Compliant Robot Navigation in Complex Indoor Environments
%I EECS Department, University of California, Berkeley
%D 2022
%8 May 4
%@ UCB/EECS-2022-36
%U http://www2.eecs.berkeley.edu/Pubs/TechRpts/2022/EECS-2022-36.html
%F Pohland:EECS-2022-36