Embodied First Exam: 1,090 embodied tasks stump coding agents, best solves just 32.9%
LINJIEFUN · x · 2026-11-09
A multi-institution team (Northwestern, CMU, Stanford, Berkeley, UT Austin, with Li Fei-Fei, Dawn Song, Pieter Abbeel, Yuke Zhu among advisors) unveiled Embodied First Exam at CoRL Late-Breaking Results: a benchmark where one command lets coding agents attempt 1,090 embodied tasks across 31 simulation suites and 44 robots, going far beyond pick-and-place to spatial puzzles, multi-robot sports, real-time non-pausing evals, whole-body control, and sim2sim AutoResearch.
It offers three modes — Privileged (full-state resets), Standard (robot API only, graded on the live run), and Realtime (the world never stops) — with four control styles: direct control, code-to-control, trained policies, and mixes.
Results are brutal: on 456 core tasks the strongest agent, GPT-6 Astra, solves only 32.9%; Claude Opus 5.5 gets 23.5%, and most agents score near zero in Realtime mode. Coding agents' first exam in the physical world is one almost nobody passes.
Related event: New Embodied Benchmark Leaves Best Agents at 32.9%(2 posts)→
More from coding & agent
- One-person site with 27 followers made $4,800 in a day, found via Claude Code revenue scan — FinanceYF5 · 2026-11-09
- The churn era: customers now use Claude to slop up replacements for overpriced SaaS — joshgonsalves_ · 2026-11-09
- Designer builds AI pipeline that finally nails map interchange generation after 2 weeks — tomjohndesign · 2026-11-09
- AI lets designers build wild ideas, but knowing the low-level stack matters more than ever — nptacek · 2026-11-09
- Gemini CLI PR fixes orphaned child processes and wrong exit codes on headless abort — duvanka-prasad · 2026-11-09
- T3 Code hands-on: one GUI orchestrates Claude Code and Codex sessions — vista8 · 2026-11-09