New Underwater Robot Uses Two Hands to Do Delicate Work

Researchers have built ULOHA, a two-armed robot that operates underwater and learns tasks by copying people. The setup has two parts: a controller rig a human moves with their own arms, and a matching robot that mirrors those movements. Once enough demonstrations are recorded, the robot's AI is trained to repeat the job on its own. The software builds on LeRobot, a free, open toolkit for teaching robots.

Water is a genuinely hard place for robots. Bubbles cloud the cameras the AI relies on. Objects drift upward or sideways instead of staying put. Water drags on every arm movement, so timing is off. The team showed the robot handling these problems: passing an object from one hand to the other, holding something with both arms at once, and intercepting a floating object mid-drift.

They also tested three popular AI approaches — ACT, Diffusion Policy, and SmolVLA, a model that sees a scene and follows instructions while moving. A separate experiment asked whether skills learned on land still work in water. The answer, under their test conditions, was yes: when the robot was trained on demonstrations from both land and water, it could perform in both.

What's the catch? This is early-stage lab research, not a product. There are no commercial robots to buy and no timeline for one. The value is that ULOHA is a shared platform other researchers can build on, so progress in this area should speed up. Down the road, two-armed underwater robots could inspect dams, offshore wind farms, ship hulls and coral reefs without putting people in the water.