Underwater Robotics: Physical AI's Toughest Test
John Leonard, an MIT professor and researcher in marine autonomy, discusses how MIT CSAIL is developing underwater robots that can map, search, and manipulate objects, including work with colleagues at the MIT Sailing Pavilion.
In this talk
- 0:00
Why underwater search is difficult
The episode opens with the unresolved search for Malaysia Airlines Flight 370 and frames underwater discovery as a challenge for physical AI. Ocean robots must cope with strong currents, low visibility, no GPS or Wi-Fi, and acoustic links that carry little data. Even a small task such as locating a mug dropped from a ship can be difficult.
- 1:30
Learning to recognize objects with sonar
Because cameras struggle in murky water, researchers use sonar, which detects objects from returning sound. The team trains camera and sonar models on paired views captured at the same time, so sonar data can be mapped to object representations learned from camera images. Building the training set required many viewpoints of objects such as oyster baskets and lobster traps, while reducing the need to label sonar images by hand.
- 3:30
Mapping and coordinating underwater robots
Leonard traces his interest in robot navigation to a factory where robots followed wires embedded in the floor. His research focuses on SLAM, which lets a robot build a map while using it to navigate. At MIT’s Marine Autonomy Lab, researchers explore cooperation among underwater vehicles, with a longer-term vision of aerial, surface, and underwater robots sharing navigation help and search data.
- 5:29
From finding objects to restoring a scene
Underwater work often still depends on divers to retrieve objects, operate valves, and make connections. Leonard describes mobile manipulation as a further challenge: a robot should understand objects well enough to move them and return a scene to its earlier arrangement. He treats this capability as a demanding test of physical AI because it requires acting despite uncertainty.
- 6:45
What ocean robotics could teach AI
Leonard envisions a robot locating and retrieving an object on the ocean floor, then restoring its surroundings. He argues that the harsh conditions underwater expose assumptions built into current AI methods, including their reliance on very large image datasets. Lessons from this setting may inform more broadly useful AI architectures.
Briefs summarise talks and interviews; timestamps open the video at that moment. Claims in a talk are the speaker's own.