RoboDojo-RC Tier 1 leaderboard

Robocurve’s rerun of six RoboDojo real-world tasks on bimanual YAM arms, with pinned rubrics and a GPT-6 Astra judge that grades each trial from three camera videos. LLM agents get 40 model calls per trial; VLAs run to a 3,600-step cap, and the benchmark README cautions that the two are not directly comparable. Scores are not comparable with RoboDojo-RealEval.

Real robot · I2RT YAM (bimanual)Report an error
65.7%Best · GPT-6.2 Astra
13Models ranked
OpenSaturation
Nov 18, 2026Numbers checked
#ModelTypeScoreDateSource
1GPT-6.2 AstraOpenAILLM controller65.7%Nov 17, 2026
2Dream-1Dream LabsWorld-action59.3%Oct 18, 2026
3Claude Fable 5.2AnthropicLLM controller57.3%Nov 9, 2026
4Gemini 4 CarbonGoogle DeepMindLLM controller54.5%Nov 5, 2026
5Gemini 4 ArgonGoogle DeepMindLLM controller48.2%Oct 14, 2026
6GPT-6 AstraOpenAILLM controller42.5%Sep 29, 2026
7GPT-6.1 SolOpenAILLM controller41.7%Oct 2, 2026
8Claude Opus 5.5AnthropicLLM controller38.2%Sep 29, 2026
9Gemini 3.8 FlashGoogle DeepMindLLM controller36.2%Oct 1, 2026
10GPT-6 SolOpenAILLM controller22.0%Sep 29, 2026
11MolmoAct2Ai2VLA18.9%Oct 16, 2026
12GPT-5.6 SolOpenAILLM controller14.9%Sep 29, 2026
13GR00T N1.7NVIDIAVLA9.5%Oct 16, 2026

Metric: Mean rubric score (0–1 partial credit per trial) over 6 real tasks, 20 trials each, graded by an LLM judge. One result per model: RoboFYI re-run, then official leaderboard, then paper, then third party. Runs that deviate from the official protocol are listed on the model page but not ranked.

Wrong or missing number? Report an error