Benchmarks/RoboColiseum

RoboColiseum

SimulationCode checkReleased Aug 2026Index weight 0.68
RoboColiseum consortiumMaintainer ↗

38 tasks in four groups, instruction following, spatial reasoning, robustness and general manipulation.

As of November 17, 2026, MXI SimWorld v0 Preview leads with 96.6, followed by HexaShell (96.5) and FF Omnimind (94.5). Official leaderboard ↗

Leaderboard

Tasks
38
Attempts / task
1
Median cost / task
$0.034
Median time / task
37 s

Score out of 100. The site reports 0 to 1.

#Model
1MXI SimWorld v0 PreviewNot counted in the index96.699.799.899.8——
2HexaShellNot counted in the index96.599.899.699.9——
3FF Omnimind94.598.698.598.5$0.034est.37 sest.4PyRUA-Lean82.899.397.984.9$17.74est.37 sest.5GO-3 Nano Preview78.889.174.682.8$0.034est.37 sest.
6Prime-0Not counted in the index77.589.667.986.5——
7HiDream-O1-EmbodiedNot counted in the index76.989.174.479.8——
8Claude Opus 5.576.396.682.579.3$6.17est.37 sest.9GPT-6 Astra76.099.378.380.4$17.74est.32 min 30 s10GPT-6.1 Sol70.995.577.770.4$2.15est.37 sest.
11DM0.5Not counted in the index70.884.461.573.4——
12Gemini 4 Argon66.885.969.473.4$3.82est.37 sest.13GigaBrain-0.765.081.747.368.0$0.034est.37 sest.
14WNDFCNot counted in the index59.871.443.864.6——
15Xiaomi-Robotics-159.477.034.665.8$0.034est.37 sest.16OpenWAM-α58.468.450.458.8$0.034est.37 sest.
17pi0.5_refineNot counted in the index57.975.539.859.3——
18π0.557.474.635.661.3$0.034est.37 sest.
19kp-robotNot counted in the index56.773.431.761.4——
20ACoT-VLANot counted in the index56.375.739.762.2——
21LingBot-VLA v255.870.842.063.3$0.034est.37 sest.22DeepSeek V4.1 Flash55.175.543.568.1$0.277est.37 sest.
23RoboMan-Pi05Not counted in the index53.673.228.853.7——
24Dream-147.267.029.859.2$0.031est.37 sest.25GR00T N1.746.864.624.953.8$0.034est.37 sest.26InternVLA-A1.546.765.027.457.5$0.034est.37 sest.27Xiaomi-Robotics-043.064.622.354.6$0.034est.37 sest.28Muse Wonder 1.042.358.631.951.6$1.52est.37 sest.
29Dream-1 zero-shotNot counted in the index33.555.315.942.0——
30π028.936.813.031.3$0.034est.37 sest.

Tasks

38 tasks. Hover a task to play its clip (37 of 38).

38 tasks · 37 with clips

Graded examples

Pick block by colour · ordered checkers

StepCheckCredit
1 · FollowGripper enters a 10 × 20 × 10 cm box around the block within 200 steps0.5
2 · PickUpOnGripperBlock rises > 2 cm and stays within 0.2 m of the gripper0.5

Reaching without lifting earns 0.5. Per-task results go only to the submitter; the board shows track means.

Methodology

Scoring
Code check
Time cap / attempt
37 s
Control rate
—
Models in index
19 of 30

Scoring. Ordered checkers in code, 0 / 0.5 / 1 per step.

Who runs it. AgiBot runs every entry on its platform, and also enters its own model.

Robots. Simulated robots in Genie Sim 3.0.

Note. AgiBot's GO-3 Nano Preview ranks first. Episodes per task are not published.

Time per attempt. Published or measured for 1 of 19 models. Estimated otherwise: time cap 37 s × (1 − success/2).

Cost per attempt. Not published. Estimated otherwise: Dream Labs API, $3 per robot-hour × time; GPT-6 Astra's median published cost per attempt on other boards; GPT-6.1 Sol's median published cost per attempt on other boards; Gemini 4 Argon's median published cost per attempt on other boards.