Benchmarks/RoboColosseum FT

RoboColosseum FT

Real robotsStaged rubricReleased Nov 2026Index weight 0.68
RoboColosseum teamMaintainer ↗

20 tasks, five per robot, from stacking and drawers to towel folding and loading a washing machine.

As of November 17, 2026, π0.5 leads with 73.1, followed by MolmoAct2 (60.5) and GR00T N1.7 (55.0). Official leaderboard ↗ · Paper ↗

Leaderboard

Tasks
20
Attempts / task
25
Median cost / task
$0.043
Median time / task
47 s

Mean progress score (%) over five tasks per robot, 25 attempts per task after fine-tuning on the same 100 demonstrations.

Tasks

20 tasks. Hover a task to play its clip (20 of 20).

20 tasks · 20 with clips

Methodology

Scoring
Staged rubric
Time cap / attempt
65 s
Control rate
—
Models in index
5 of 5

Scoring. Judges score each run against criteria levels 0 to K per task.

Who runs it. The same operator network runs every trial.

Robots. Franka (DROID setup), two-arm YAM, SO-101, Unitree G1.

Note. Every model is fine-tuned on the same 100 demonstrations per task, then gets 25 attempts. Scored by partial progress.

Time per attempt. Not published. Estimated otherwise: time cap 65 s × (1 − success/2).

Cost per attempt. Not published. Estimated otherwise: one H100 at $3.29 per hour (Lambda on-demand, Oct 2026) × time. GPU price source ↗