← All benchmarks

EmbodiedLab (T2 success, distilled skills)

Agents · 2026-10-26

EmbodiedLab's diagnostic level T2: T1 plus a workspace library of locate, grasp, carry, place, press, push and slide skills distilled from the GPT and Claude agents' successful runs on the same 30 experiments. The paper uses it only to measure how much failure lies in acting, keeps it out of the released protocol, and runs it only for agents that donated no skills, because a T2 run partly reuses known solutions; it is therefore a diagnostic. The score is the T2 success rate, averaged over each experiment's runs and then over the experiments, in percent.

Top models (higher is better)