Benchmarks

Benchmarks

The simulation suites we run a policy against, grouped by the arm layout it is scored on.

Single-armone arm, one gripper

One arm and a gripper on a fixed base — the layout most tabletop manipulation policies are trained and scored on.

single-arm
LIBERO
libero
130 tasks · Franka Panda · four suites — spatial, object, goal and long-horizon · language-conditioned
Bi-manualtwo arms in coordination

Two arms that have to be coordinated — dual-arm rigs and humanoid upper bodies, where the timing between the hands is part of the task.

bi-manual
RoboCasa GR1
robocasa-gr1
24 tabletop tasks · GR-1 humanoid with Fourier hands · 18 pick-and-place, 6 articulated · hand–arm coordination
bi-manual
RoboTwin 2.0
robotwin-2
50 dual-arm tasks · five embodiments · strong domain randomization — clutter, lighting, background, height, language