SEAR: first large-scale controlled study of self-evolving robot agents, 7 LLMs, 330 tasks
xiye_nlp · x · 2026-10-16
SEAR (arXiv:2610.18792) is billed as the first large-scale controlled study of self-evolving robot agents, asking whether rising success rates reflect real learning or just more attempts. Agents improve robots in three modes: code-as-policy, agent-as-policy (live control loop), and auto-research (fine-tuning a VLA). The benchmark spans 7 frontier LLMs, 330 simulated and 5 real-robot tasks, 4+ years of agent runtime, with matched controls and a sealed judge on unseen scenes. Findings: learning curves depend on time budget and task difficulty relative to model capability; control code and shared routines transfer across tasks and simulators while task-specific solutions don't, with a real-robot gap between agent practice and novel human-designed scenes; different models evolve better in different ways.
More from Embodied
- Viral meme nails AI's irony: it can write symphonies but can't do the dishes — briannekimmel · 2026-10-16
- WEDGE turns navigation agents' own failures into targeted training worlds — mohitban47 · 2026-10-16
- Knödel Watch: animated vocabulary cards on ESP32-S3, open-sourced 4 days after unboxing — alexcovo_eth · 2026-10-16
- SlipSense fuses pressure and vibration for 23ms zero-shot robotic slip detection — micoolcho · 2026-10-16
- Stanford's Fei-Fei Li team open-sources OpenWAM, a unified world-action model framework for robots — jiqizhixin · 2026-10-16
- Robot folding leaderboard: GPT-6 Astra leads at 1349 Elo, Muse Wonder trails at 1034 — qinzytech · 2026-10-16