Co-Chairs: Maani Ghaffari Jadidi and X. Jessie Yang


Abstract:
Socially Assistive Robots (SARs) help humans through social interactions. As robots are increasingly integrated into human spaces, a key challenge is enabling them to navigate semi-structured environments. This thesis addresses robot behavior and motion planning in dynamic, human-centric scenarios, using a tour guide robot as a test case.

The first part focuses on tour planning with a shared map created collaboratively by Providers (managers) and Robots to guide Clients (visitors). The planner dynamically adapts routes based on constraints, highlighting the importance of shared maps in human-robot tasks.

The second part explores low-level motion planning. We start with crowd navigation, designing an agent to move through groups without disrupting the human flow. We then tackle narrow crossings, using Smooth Maximum Entropy Deep Inverse Reinforcement Learning (S-MEDIRL), the robot learns from raw data to yield and avoid deadlock without relying on handcrafted heuristics (see attached gif).

We then evaluate Foundation Models in Socially Assistive settings, demonstrating a robot as a greeter and tour guide at the University of Michigan Museum of Art (UMMA). The second part explores incorporating feedback from Vision and Language Models to distinguish between qualitatively good and bad social behaviors. We evaluate the efficacy of these generalized models in supervising and improving socially aware navigation, specifically in the narrow crossing scenario.

In conclusion, this thesis develops motion and behavior planning modules to enable socially aware robot navigation in human environments.

Zoom passcode: fetch