Mind2Act World Benchmark Drops Top Agent to 12% on Composition
TL;DR
- The strongest agent, GPT-6 Astra, scored 52.7% on Mind and 39.0% on Act episodes but only 12.0% on composed Mind2Act episodes.
- Eight of the 30 tasks were completed by zero systems across all 56,700 evaluation episodes.
- Best end-to-end policy InternW0 reached 8.7% overall, and a single task, gate passage (DR3), accounts for 36-70% of end-to-end successes.
Across 56,700 simulated episodes on a new embodied-intelligence benchmark, eight of the 30 tasks were completed by zero systems. Not once.
The benchmark is Mind2Act World, from a Tsinghua-led team with co-authors at Shenzhen University and Velabot AI. It bundles 30 manipulation tasks into 180 configurations, split across three views: Mind (history-based reasoning), Act (precise and timely execution), and Mind2Act, where action outcomes guide subsequent decisions in closed loop. The robot is a bimanual ARX X5 rig in Isaac Sim.
"Even the strongest agent succeeds in 52.7% of Mind and 39.0% of Act episodes but in only 12.0% of Mind2Act episodes," the paper reports. The strongest agent is GPT-6 Astra. If the two capabilities were independent, composition would predict 20.6%; the observed 12.0% comes in roughly 58% below that.
End-to-end policies fared worse. The top vision-language-action model, InternW0, hit 8.7% overall and 3.0% on Mind2Act. "Gate passage (DR3) accounts for 36-70% of all end-to-end policy successes," meaning most systems cluster their wins on a handful of practiced motions. The eight tasks nobody completed include workpiece tracking, tray carrying with loose balls, stamp angle control, and a domino chain to bell.
Difficulty scales punishingly on the composition view: GPT-6 Astra scored 27.3% on Easy Mind2Act, 8.5% on Medium, and 4.2% on Hard. The authors point to "the composition of reasoning and execution through environmental feedback as the main open problem."
The paper lands in a dense week on our robotics tracker, which logged 218 stories in the past 90 days, including AFFECT and ZeroVLN earlier the same day.
Originally reported by huggingface.co
Read the original article →Original headline: Mind2Act World Benchmark Shows Top Robots Succeed on 52.7% of Mind Tasks But Only 12% When Composed