Robota-Embed: MIT team's unified VLA embedding beats Gemini Embedding 2 by 22 points, lets robots learn from past experience
songhan_mit · x · 2026-10-17
A team from MIT, Berkeley and partners introduced Robota-Embed, described as the first unified vision-language-action (VLA) embedding model, jointly encoding task, execution stage and future continuations so robots can retrieve useful context from recorded experience.
Key results:
- Retrieval: 47.33% clip-level R@1, beating action-free Gemini Embedding 2 by 22.09 points (81.77% episode-level, +33.47 points)
- Robot learning: Embed2ICL uses retrieved experience for policy conditioning, raising success from 47.0%→56.3% on DROID-Sim and 86.4%→92.8% on LIBERO's hard subset, with gains on real robots
- Self-evolution: Embed2Evolve grows task-local memory from execution feedback — 24%→68% success on ManiSkill while using 43.7% fewer policy tokens than a direct GPT-6 Astra baseline
Trained on 1.7M cleaned anchors, with Robota-Embed-Bench (4,000 queries, 8,000-entry gallery). Authors include Yuke Zhu, Trevor Darrell, Linxi "Jim" Fan, Song Han and Han Cai.
Related event: Robota-Embed: Unified VLA Embeddings Enable Robot In-Context Learning(2 posts)→
More from Embodied
- Harness Pro: a $299 desk screen to watch and direct AI agents on every machine — dee_hw · 2026-10-17
- X-Reset eases RL exploration for dexterous hands by resetting to human hand-object states — weichiuma · 2026-10-17
- Princeton's LabOS, Featured in Nature: An AI Co-Scientist for Hypothesis and Experiment Design — MengdiWang10 · 2026-10-17
- Founder's China trip: robotics has no single hub, and robot firms work less hard — DominiqueCAPaul · 2026-10-17
- Standard Bots CEO: robots are at their '1973 moment'—and his plan to win US manufacturing — Rewkang · 2026-10-17
- Google Gemma hardware hackathon at AGI House offers $45k prizes, $15k for LiteRT track — agihouse_org · 2026-10-17