PhD Thesis Defense: Michael Everett

In this talk

1. 0:00

   Bringing robots into everyday spaces

   Everett argues that robots need to navigate without prior maps, account for people and other moving agents, and remain safe without becoming so cautious that they cannot do their jobs. He frames these as perception, planning, and safety challenges, using delivery, hospital, and household robots as examples.

2. 5:07

   Using context to find a goal

   For tasks such as reaching a home's front door, Everett's Deep Cost-to-Go method predicts likely routes from a partial semantic map. It learns from other houses, then guides exploration in places it has not seen before. In tests, it improved path efficiency over frontier-based exploration and transferred from aerial house maps to a 3D simulation using a forward-facing camera.

3. 16:22

   Scaling collision avoidance

   Everett describes learning collision-avoidance policies in simulation and deploying them on ground robots and flying robots. An LSTM processes a changing number of nearby agents, helping the policy handle crowded settings and scenarios with more agents than it encountered during training. The experiments included a mobile robot among pedestrians and four hexarotors moving toward assigned goals.

4. 31:35

   Making learned decisions robust

   Everett examines how small sensor changes can disrupt learned policies, including adversarial inputs that alter an autonomous vehicle's lane classification. His approach uses certified bounds on reinforcement-learning action values to choose an action that accounts for a range of possible states. In collision-avoidance simulations, greater robustness reduced crashes, but excessive caution lowered reward by making agents reluctant to reach their goals.

5. 44:27

   Future directions

   Everett suggests learning multiple possible routes to a goal, combining robot motion models with learned models of human behavior, and adjusting robustness online as uncertainty changes. These directions aim to make autonomous systems more capable in unfamiliar, crowded, and imperfectly sensed environments.