Ahalya Prabhakar on Coverage-Based Control for Robot Communication and Learning |Toronto AIR Seminar

In this talk

1. 4:06

   Ergodic control turns tasks into spatial goals

   Prabhakar describes a method that steers robots to match how often they visit different parts of a workspace to a target spatial distribution. It optimizes the trajectory’s overall statistics, rather than assigning a robot a precise location at a precise time. She uses Fourier coefficients to compare the desired distribution with the robot’s observed motion, and incorporates that measure into model predictive control.

2. 7:51

   Swarm operators specify regions of interest

   For human-swarm coordination, an operator shades areas on a tablet to indicate where the swarm should spend more or less time. Agents share compact trajectory statistics and adjust their own actions to meet the group’s goal, even as priorities or swarm size change. In simulations and DARPA OFFSET field tests, this approach supported groups of up to 50 agents; when agents were disabled, the others redistributed to cover the requested regions.

3. 18:18

   Decentralized coverage helps handle field failures

   Prabhakar reports that swarms could adapt to signal loss, hardware failures, and agents getting stuck, without an operator assigning new routes. A disconnected rover could continue working toward the shared coverage goal using its onboard specification. Agents exchanged Fourier coefficients, keeping communication tied to the number of coefficients rather than the length of their travel histories.

4. 22:22

   Robots gather data to learn sensor models

   Prabhakar uses conditional variational autoencoders to learn how sensor readings relate to robot state, then steers exploration toward regions with high decoder output entropy. In tests with a camera viewing a cube and sphere, targeted sampling produced useful object models with less energy than random sampling; information maximization tended to favor one object. She also demonstrated learning the appearance of a rubber duck and a simulated electrosensing model without supplying the robot with object or sensor physics.

5. 41:47

   Applications and limits depend on the task

   Prabhakar says the approach has been tested in real-world search and communication tasks, but intuitive spatial specifications become harder to use in higher-dimensional spaces. A coverage goal assumes the requested behavior is feasible, and coverage alone cannot guarantee that a robot avoids a dangerous region; that requires a barrier function or another safety method. In DARPA OFFSET, the team used coverage maps to share likely target locations and mark areas containing adversarial agents, while coping with unreliable rover communications.