← All benchmarks

EmbodiedLab (T0 success)

Agents · 2026-10-26

EmbodiedLab, from HexaFuture and collaborators, has 30 quantitative science experiments in mechanics, electricity, optics, thermal physics and chemistry on 14 apparatus platforms. A coding agent operates a Franka arm with a two-finger gripper and three cameras in an Isaac Sim laboratory through a Python client, chooses its own procedure and reads the instruments from the camera images; the hidden system parameters are redrawn for every instance. A run succeeds only if every submitted quantity is within its instrument-set tolerance (0.2% to 5%), the cited readings were recorded in the same episode and the bench is restored. At the headline T0 level the client supplies only the camera images, the robot state, seven-joint position targets and gripper commands. The score is the T0 success rate, averaged over each experiment's runs and then over the 30 experiments, in percent.

Top models (higher is better)