Researchers have developed NereidVLA, a novel vision-language-action model designed for underwater navigation. The model first learns marine spatial knowledge from approximately two million examples through its Sea2Know component, which focuses on target grounding and spatial relations without direct robot motion data. This pre-trained model is then fine-tuned with expert trajectories from simulated underwater vehicles. NereidVLA translates RGB histories and language instructions into motion commands for closed-loop navigation, achieving a 78.75% success rate in laboratory pool trials, significantly outperforming the InternVLA-N1 model. AI
IMPACT This model could significantly improve the autonomy and capability of underwater robots for exploration and tasks.
RANK_REASON The cluster describes a new model and benchmark results published on arXiv, fitting the research category. [lever_c_demoted from research: ic=1 ai=1.0]
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →