Underwater VLA policies beat baseline; ACT beats SmolVLA at sea
Original title: Embodied at Sea: Imitation Learning for Multi-Task Underwater Manipulation
A 14-model DynamicVLA study and an ROV case study at a Malta wreck also open-source the BLUEMBODIED KIT.
Researchers collected 1,616 underwater manipulation episodes in calm and turbulent water and trained 14 DynamicVLA models with different proprioception setups. On 1,260 gradable pool rollouts across three tasks, 10 models beat the baseline; the best reached a 93.3% overall success rate, including partial successes. A separate ROV case study on grasping and mooring around the WWII wreck Coralita X127 and at Manoel Island Marina, Malta, found an Action Chunking Transformer far ahead of SmolVLA. The BLUEMBODIED KIT for data collection, training, and deployment is open sourced.
- Case study used 748 episodes and 508 gradable on-hardware rollouts.
- ACT pooled success was 47.9% versus 7.9% for SmolVLA.
- Best archaeological-site task reached a 66.7% ACT success rate.
- Field tests used an ROV at Coralita X127 and Manoel Island Marina.
- Products
- BLUEMBODIED KIT
- Countries
- Malta
- AI models
- DynamicVLASmolVLA
From the original article
1 min read at arxiv.orgBehavior Cloning has demonstrated promising generalization capabilities for robotic manipulation in terrestrial environments. However, marine visual distortions and dynamic currents create a severe domain shift that remains a key bottleneck in adapting generalist architectures to underwater field manipulation.
By Greta Perez-Haiek · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.