评测成绩#

RPent Leaderboard

LIBERO-PRO

方法总体Spatial TaskSpatial SwapObject TaskObject SwapGoal TaskGoal SwapLong TaskLong Swap
RPent / GPT-6 Astra[4]Codex · low · reasoning 92.63% (741/800) 100% (100/100) 98% (98/100) 100% (100/100) 99% (99/100) 88% (88/100) 99% (99/100) 85% (85/100) 72% (72/100)
RPent / DeepSeek V4.1 FlashCodex · xhigh · reasoning 90.00% (720/800) 98.00% (98/100) 98.00% (98/100) 100.00% (100/100) 100.00% (100/100) 90.00% (90/100) 93.00% (93/100) 74.00% (74/100) 67.00% (67/100)
RPent Flash Mode[2]Molmo2-8B · 视觉定位 72.63% (581/800) 79.00% (79/100) 70.00% (70/100) 86.00% (86/100) 93.00% (93/100) 74.00% (74/100) 65.00% (65/100) 60.00% (60/100) 54.00% (54/100)
RPent / Qwen3.6 27Bno-reasoning 70.63% (565/800) 82.00% (82/100) 78.00% (78/100) 83.00% (83/100) 84.00% (84/100) 68.00% (68/100) 68.00% (68/100) 61.00% (61/100) 41.00% (41/100)
RPent / GPT-5.5Codex · xhigh · reasoning 75.13% (已评测: 800) 81.0% (已评测: 100) 69.0% (已评测: 100) 94.0% (已评测: 100) 91.0% (已评测: 100) 75.0% (已评测: 100) 66.0% (已评测: 100) 70.00% (70/100) 55.00% (55/100)
RPent / GPT-6 Motor Only[3]Codex · low · reasoning — — — — — — — 38.0% (38/100) —
π0 0.3% 0.0% 0.0% 0.0% 2.0% 0.0% 0.0% 0.0% 0.0%
π_RLinf 50.0% 42.0% (已评测: 100) 59.0% (已评测: 100) 71.0% (已评测: 100) 78.0% (已评测: 100) 45.0% (已评测: 100) 42.0% (已评测: 100) 49.0% (已评测: 100) 14.0% (已评测: 100)
AtomVLA 6.3% 1.0% 16.0% 0.0% 10.0% 11.0% 2.0% 9.0% 1.0%
π0.5 11.0% 1.0% 20.0% 1.0% 17.0% 2.0% 38.0% 1.0% 8.0%
MolmoAct 1.5% 0.0% 0.0% 0.0% 6.0% 0.0% 0.0% 6.0% 0.0%
X-VLA 3.8% 0.0% 0.0% 8.0% 2.0% 9.0% 1.0% 10.0% 0.0%
Cap-X — 14.0% 12.0% 18.0% 22.0% 17.0% 26.0% — —
RATS — 31.0% 29.0% 63.0% 61.0% 36.0% 43.0% — —
RPent / Opus-4.7Claude Code · max · reasoning 82.4% (已评测: 800) — — — — — — — —
ASPIRE[1] 61.36% 60.0% (已评测: 500) 51.0% (已评测: 500) 95.0% (已评测: 500) 98.0% (已评测: 500) 45.0% (已评测: 500) 81.0% (已评测: 500) 38.3% (已评测: 500) 22.6% (已评测: 500)

* [1] ASPIRE:Long Task 和 Long Swap 使用 LIBERO-90 技能库进行 zero-shot 迁移。

* [3] RPent / GPT-6 Motor Only 为纯电机控制方法,通过 execute_action 直接输出末端位姿增量与夹爪指令,不调用 VLA / primitive,不加载 memory。

* [4] GPT-6 Astra(使用 memory 的配置):Long Task/Swap 与其余六套件使用各自探索后冻结的 memory 文件快照,评测期间不更新。Overall 合并两个不重叠批次:Long 157/200,加上其余套件 584/600,得到 741/800(92.63%);并非全部回合共享同一份 memory 快照。

LIBERO

方法总体SpatialObjectGoalLong
RPent / Opus-4.7Claude Code · max · reasoning 96.0% (384/400) 97.0% (已评测: 100) 100.0% (已评测: 100) 94.0% (已评测: 100) 93.0% (已评测: 100)
OpenVLA 76.5% 84.7% 88.4% 79.2% 53.7%
NORA 79.5% 85.6% 89.4% 80.0% 63.0%
π0 94.2% 96.8% 98.8% 95.8% 85.2%
π_RLinf 95.3% 99.0% (已评测: 100) 96.0% (已评测: 100) 97.0% (已评测: 100) 89.0% (已评测: 100)
AtomVLA 97.0% 96.4% 99.6% 97.6% 94.4%

RoboCasa365 · Target50

方法总体Atomic-SeenComposite-SeenComposite-Unseen
RPent / GPT-6 AstraCodex · low · reasoning 59.20% (已评测: 340) 87.78% (已评测: 180) 43.75% (已评测: 80) 42.50% (已评测: 80)
RPent / GPT-5.5Codex · xhigh · reasoning 57.1% (已评测: 340) 92.0% (已评测: 180) 61.0% (已评测: 80) 13.8% (已评测: 80)
RPent / Opus-4.7Claude Code · max · reasoning 48.6% (已评测: 340) 79.4% (已评测: 180) 47.5% (已评测: 80) 15.0% (已评测: 80)
π0 14.8% 34.6% 6.1% 1.1%
π0.5 16.9% 39.6% 7.1% 1.2%
RLDX-1 30.0% 60.0% 21.3% 5.0%
WorldDreamer 35.3% 66.3% 26.7% 9.0%
Xiaomi-Robotics-1 57.4% 80.2% 57.1% 32.1%

RoboTwin

方法C2R
RPent / GPT-5.5Codex · xhigh · reasoning 62.4% (156/250)
RPent / Opus-4.7Claude Code · max · reasoning 58.4% (已评测: 250)
π0.5 47.9%
GR00T-N1.7 20.7%
StarVLA 10.6%
LingBot-VLA 50.4%

* Leaderboard 中所有 RPent 方法均展示 2026.09 测试结果。

* LIBERO / LIBERO-PRO:所有使用 VLA 的 RPent 方法均采用与 π_RLinf 基线相同的 RLinf-Pi05-LIBERO-130-fullshot-SFT checkpoint。该模型由 RLinf 发布,已经过 LIBERO-130 监督微调;RPent 未针对 LIBERO-PRO 额外微调,评测期间权重冻结。Motor Only 不使用 VLA。

* RoboCasa365:RPent 使用 RLWRLD 官方发布的 RLDX-1-FT-RC365,而非 RLinf checkpoint。该模型已针对 RoboCasa365 微调;RPent 未额外微调其权重,评测期间保持冻结。

* RoboTwin:RPent 使用 RLinf 发布的 LingBot-VLA-RoboTwin-EEF-ckpt1500,该 LingBot-VLA checkpoint 已经过 RoboTwin 数据监督微调。RPent 与直接 LingBot-VLA 基线使用相同 checkpoint,评测期间均冻结权重。

* 各方法的训练数据、探索 / memory、推理预算及结果选择方式可能不同;外部方法按原来源报告,同 checkpoint 对照见各环境的 VLA 说明。

* [2] RPent Flash Mode 使用直接下载的、官方公开的 GPT-5.5 explore memory,Molmo2-8B 用于视觉定位,动作执行使用与 π_RLinf 相同的 VLA。展示成绩采用 s0-s9 中的最佳评测结果。

RPent Leaderboard