Time & Token Costs#

RPent Leaderboard

LIBERO-PRO
MethodMean time / episode (s)Total output tokensMean input + output tokens / episode
RPent / GPT-6 Astra[4]Codex · low · reasoning412.143,364,938—
RPent / DeepSeek V4.1 Flash[5]Codex · xhigh · reasoning963.25—5,229,712
RPent / DeepSeek V4.1 Flash[6]Codex · xhigh · reasoning632.65—4,309,012
RPent / GPT-5.6 SolCodex · xhigh · reasoning529.233,635,426—
RPent / GPT-5.6 SolCodex · no-reasoning344.172,569,676—
RPent Flash ModeMolmo2-8B · visual localization60.190—
RPent / Qwen3.6 27Bno-reasoning626.303,270,754—

* [4] GPT-6 Astra (memory-enabled configuration): Long Task/Swap and the other six suites use separate memory-file snapshots frozen after their respective exploration phases, with no updates during evaluation. Overall combines two non-overlapping batches: Long 157/200 plus the other suites 584/600, giving 741/800 (92.63%); the 800 episodes do not share a single memory snapshot.

* [5] DeepSeek V4.1 Flash evaluated with memory explored by GPT-6 Astra.

* [6] DeepSeek V4.1 Flash evaluated with its own seed-0 exploration memory.

RoboCasa365 · Target50
MethodMean time / episode (s)Total output tokens
RPent / GPT-6 AstraCodex · high · reasoning1,168.955,894,436
RPent / GPT-5.6 SolCodex · xhigh · reasoning1,197.565,258,391
RPent / GPT-5.6 SolCodex · no-reasoning1,013.214,792,104

RoboTwin
MethodMean time / episode (s)Total output tokens
RPent / GPT-6 AstraCodex · low · reasoning1,078.802,763,641
RPent / GPT-5.6 SolCodex · xhigh · reasoning1,828.704,763,915
RPent / GPT-5.6 SolCodex · no-reasoning957.402,456,401

* RPent results on this leaderboard come from different evaluation dates; each source note identifies its run.

* LIBERO / LIBERO-PRO: all VLA-based RPent methods use the same RLinf-Pi05-LIBERO-130-fullshot-SFT checkpoint, released by RLinf, as the pi_RLinf baseline. This checkpoint was supervised-fine-tuned on LIBERO-130; RPent does not further fine-tune it for LIBERO-PRO, and its weights remain frozen during evaluation. Motor Only does not use a VLA.

* RoboCasa365: RPent uses RLDX-1-FT-RC365 released by RLWRLD, not an RLinf checkpoint. It was fine-tuned on RoboCasa365; RPent does not further fine-tune its weights and keeps them frozen during evaluation.

* RoboTwin: RPent uses LingBot-VLA-RoboTwin-EEF-ckpt1500 released by RLinf, a LingBot-VLA checkpoint supervised-fine-tuned on RoboTwin data. RPent and the direct LingBot-VLA baseline use the same checkpoint with frozen weights during evaluation.

* Methods may differ in training data, exploration / memory, inference budgets, and result selection. External results follow their original sources; the environment-specific VLA notes identify same-checkpoint comparisons.

* [2] RPent Flash Mode uses directly downloaded, officially released GPT-5.5 exploration memory, Molmo2-8B for visual localization, and the same VLA as pi_RLinf for action execution. The displayed scores are the best evaluation results from s0-s9.

RPent Leaderboard