Skip to content

Benchmarks

Benchmark Simulator Tasks Triaged
BEHAVIOR-1K OmniGibson (NVIDIA Isaac Sim) 100 0/100
DexToolBench MuJoCo 3.3.7 (our port; upstream runs Isaac Sim / Isaac Gym) 24 21/24
HumanoidBench MuJoCo 3 (HumanoidBench's gymnasium environments) 9 + 15 excluded 9/9
KinDER MuJoCo 3.3 (KinDER's Dynamic3D environments); PyBullet for two kinematic tasks now excluded 4 + 7 excluded 4/4
MetaWorld+ MuJoCo 3.3.0 (Meta-World v3) 50 50/50
MolmoSpaces MuJoCo 3.5.0 (molmo-spaces 0.2.9) 8 8/8
MuJoCo Playground (manipulation) MuJoCo 3.3.7 (plain MuJoCo on the CPU; upstream trains in MJX / MuJoCo Warp) 10 4/10
RoboCasa-GR1 robosuite @ 51cc017 (1.5.1, mink whole-body IK) / MuJoCo 3.2.6 (RoboCasa GR1 tabletop) 24 24/24
RoboCasa robosuite 1.5.1 / MuJoCo 3.2.6 (RoboCasa v0.2) 29 29/29
RoboCasa365 robosuite 1.5.2 / MuJoCo 3.3.1 (RoboCasa365) 50 50/50
RoboLab Isaac Lab 2.3.2 / Isaac Sim 5.1 (NVIDIA) 120 0/120
RoboPaint ManiSkill 3.0.1 / SAPIEN 3.0.3 + Spline-FRIDA stroke model 82 + 9 excluded 0/82
RoboTwin 2.0 SAPIEN 3 (PhysX) + CuRobo 50 50/50
RoboWits Genesis 0.4.7 (gs_gym) 29 + 1 excluded 0/29
VLABench MuJoCo 3.2.2 + dm_control 1.0.22 (VLABench) 36 36/36
Each benchmark has the same three pages:
Page Contents
Tasks Every task as a row with its demo playable in place — the landing page
Triage board Keep/drop decisions and what is still open
Capability coverage What the suite exercises, and the thin spots

Adding another benchmark is a YAML file plus task pages; every view picks it up automatically. See Adding a benchmark.