AQUAACT: A Joystick-Level Underwater Teleoperation Dataset and Closed-Loop Benchmark

Key points

• Underwater robots support a growing range of work, from inspecting offshore structures and pipelines to surveying reefs and accompanying divers.1
• On land, vision-language-action (VLA) models already let robots carry out tasks autonomously from language instructions, which makes them a natural candidate for taking the pilot's place.1
• Underwater, however, there is little data to train such models.1
• A closed-loop benchmark of seven policies in simulation, from imitation-learning baselines to VLA models, each trained with its default recipe, shows that these tasks remain far from solved: no policy completes more than a fifth of the episodes.1

Sources (1)

• [1]AQUAACT: A Joystick-Level Underwater Teleoperation Dataset and Closed-Loop Benchmark
  arXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 18, 09:07 PM

  We therefore record the pilots and introduce AQUAACT, to our knowledge the first publicly available language-annotated underwater dataset pairing camera images with human joystick commands.

  Underwater robots support a growing range of work, from inspecting offshore structures and pipelines to surveying reefs and accompanying divers.

Extractive summary: sentences quoted from the sources.