Driven by 1,616 demonstrations in calm and agitated water, vision-language-action models reach 93.3% success in the basin. On a wreck, an ACT policy succeeds 66.7% of a task.
Visual distortions of the marine environment and currents make it difficult to transfer robotic manipulation methods that work well on land underwater. A team collected 1,616 episodes of underwater manipulation, in both calm and turbulent water, to train 14 variants of a generalist vision-language-action model (DynamicVLA) with different proprioceptive information.
From the basin to the wreck
Out of three tasks and 1,260 tests evaluated in the pool, ten variants exceeded the benchmark; the best achieved 93.3% success, including partial successes. The researchers then tested two policies on an ROV in real conditions, around the World War II wreck Coralita X127 and in the marina of Manoel Island, Malta, for marine archaeology and conservation uses (seizure and mooring). Out of 508 trials, the specialized ACT policy clearly outperformed the generalist SmolVLA model, with 47.9% success rate versus 7.9%, and up to 66.7% over the best task conducted at the archaeological site.
The collection, training and deployment kit, BLUEMBODIED KIT, is published as open source.
Read the full article on the original source
Go to the source →