Fall 2026 GRASP on Robotics: Stephanie Gil, Harvard University
Stephanie Gil, a robotics researcher at Harvard University, explains how multi-robot systems can use communication signals to build trust, resist attacks, and make decisions under uncertainty.
In this talk
- 6:44
Physical signals can help robots assess trust
Gil shows how wireless signals can reveal information about robots’ relative positions and whether agents are distinct. For example, signal fingerprints measured with commodity Wi-Fi hardware can expose a Sybil attack, where one adversary poses as multiple agents. These physical measurements provide evidence that can help check the data agents send.
- 13:27
Trust measurements strengthen robot consensus
In consensus, robots share values such as heading directions and adjust their behavior to agree. Gil’s method uses noisy trust measurements, called alpha values, to change how much weight each agent gives its neighbors. When measurements distinguish legitimate links from malicious ones on average, the team can eventually classify them and reach agreement even with many malicious agents, though those agents can still shift the final value.
- 33:24
The same approach extends to distributed optimization
Gil applies trust-weighted updates to distributed optimization and hypothesis testing, where agents combine information while adversaries may try to distort the result. In optimization, trusted neighborhoods can recover the legitimate agents’ target, and a newer approach adds projection steps to recover faster geometric convergence. The talk also describes tests with 50 legitimate agents and 100 malicious agents.
- 42:04
Robots route to sperm whales using noisy observations
For natural uncertainty, Gil describes routing a drone to meet sperm whales at sea, where a surfacing opportunity lasts about 10 minutes per hour. A boat’s acoustic array estimates whale direction from vocalizations, and the system uses those noisy bearings to plan the drone’s route. In an experiment, the drone autonomously came within about 187 meters of a whale before a human operator took over.
- 51:12
Future work connects trust with learning and planning
Gil wants to use physical signals to assess whether data is reliable for machine learning tasks and robot control. Her group is also combining sequential decision-making with adversaries, and exploring whether information in whale vocalizations can improve state estimation and planning.
Briefs summarise talks and interviews; timestamps open the video at that moment. Claims in a talk are the speaker's own.