Robota-Embed aims to help robots retrieve and learn from past experiences

Robota-Embed is described in its introductory post as a vision-language-action embedding model for retrieving past robot experiences. The post reports a 47.33% R@1 retrieval score, 22.09 points above Gemini Embedding 2, and says robot-learning success rates rose from 47.0% to 56.3% on DROID-Sim and from 86.4% to 92.8% on LIBERO. On ManiSkill, it claims task success rose from 24% to 68% with 43.7% fewer policy tokens than a direct GPT-6 Astra baseline.