Embodied First Exam Launches: One Line of Code to Test Coding Agents on 1,090 Embodied Tasks

Manling Li's team, together with researchers including Guanya Shi, has released the Embodied First Exam benchmark: with a single line of code, a coding agent can run through 1090 embodied tasks across 31 simulation suites and 44 robots. The benchmark answers one core question—how would coding agents that have passed many exams in the digital world perform on their first exam in the physical world, and how far are we from general-purpose robots? The project has opened its website, technical report, code, and Discord, and is openly inviting community participation.

Confirmed

• The benchmark covers 31 simulation suites, 44 robots, and 1090 embodied tasks, going far beyond traditional pick-and-place to include spatial constraint puzzles, multi-robot collaborative Olympic-style events, real-time tasks, and more
• According to the authors' launch post, this is agent-native evaluation infrastructure specifically designed to test coding agents' core robotics capabilities: reasoning and planning, contact-rich manipulation, dynamics and latency, safety, whole-body motion control, competition and collaboration, and sim2sim
• The project also released a technical report, open-sourced the code, and opened a Discord community channel

Why it matters

• The authors point out that existing evaluations mostly stay at simple pick-and-place tasks and struggle to measure a coding agent's overall robotics capability; this benchmark extends evaluation to safety, dynamics, collaboration, and other dimensions closer to real robots
• The low-barrier "one line of code" design lets all kinds of general-purpose coding agents directly take exams in physical-world scenarios, potentially making it one of the standardized evaluation infrastructures for embodied AI

2026-10-20 ~ 2026-10-20 · 6 related posts

• [source] Embodied First Exam: one line of code to test coding agents on 1,090 embodied tasks — ManlingLi_ · 2026-10-20
• [source] Agent-native robotics eval infra: 1,090 tasks, 2,128 variants, 44 robots — ManlingLi_ · 2026-10-20
• Embodied First Exam launches as the first physical-world benchmark for code agents — ManlingLi_ · 2026-10-20
• Embodied First Exam: Top Model Scores Just 33% Across 1,090 Robotics Tasks for Coding Agents — ManlingLi_ · 2026-10-20
• [source] Embodied First Exam: 1,090 robotics tasks in one line of code, backed by $500K Robocurve fund — ManlingLi_ · 2026-10-20

1 near-duplicate retellings: wzenus