Benchmarks/RoboDojo-RC

RoboDojo-RC

Real robotsStaged rubricReleased Sep 2026Index weight 0.9
Robocurve, Inc.Maintainer ↗

6 RoboDojo tasks adapted, Cap Pen, Pack and Pour Fruit, Sort Objects, Stack Bowls, Stand Up Bottles, Store In Safe.

As of November 17, 2026, Dream-1 leads with 59.3, followed by Claude Fable 5.2 (57.3) and Gemini 4 Carbon (54.5). Official leaderboard ↗

Leaderboard

Tasks
6
Attempts / task
20
Median cost / task
$1.77
Median time / task
4 min 02 s

Overall vs speed

Published time per task attempt

Faster, higher scoring20.040.060.03 min 20 s8 min 20 s16 min 40 s33 min 20 sTime per task (faster ←)Overall (better ↑)Claude Fable 5.2GPT-6 AstraClaude Opus 5.5

Overall vs cost

Published cost per task attempt

Cheaper, higher scoring20.040.060.0$0.500$1.00$2.00$5.00Cost per task (cheaper ←)Overall (better ↑)Claude Fable 5.2Gemini 4 CarbonGPT-6.1 SolGemini 3.8 Flash

Progress score 0 to 100, 20 attempts per task.

Tasks

6 tasks. Hover a task to play its clip (6 of 6).

6 tasks · 6 with clips

Methodology

Scoring
Staged rubric
Time cap / attempt
—
Control rate
—
Models in index
12 of 12

Scoring. An LLM judge applies written rubric levels to each trial video.

Who runs it. Independent evaluator (Robocurve).

Robots. I2RT YAM (two-arm).

Note. Tests coding agents and robot models on the same tasks. Tasks chosen after trying GPT-6 Astra.

Time per attempt. Published or measured for 9 of 12 models. Estimated otherwise: median measured time on this board (9 models).

Cost per attempt. Published for 9 of 12 models. Estimated otherwise: Dream Labs API, $3 per robot-hour × time; one H100 at $3.29 per hour (Lambda on-demand, Oct 2026) × time. GPU price source ↗