外观
Skip to content
榜单与评测 · 公开基准跑分与开源热度
具身智能基准榜单中心 (Leaderboards)
汇总 VLA / 具身操作策略在 LIBERO、SimplerEnv、CALVIN 等学术基准上的公开论文成绩，以及 GitHub 机器人开源项目星标榜与 Hugging Face 具身模型/数据集下载榜。所有分数均注明原始论文或报告出处。
Franka 桌面操作四套件（各 10 个任务）成功率，单位 %。 avg 是 Spatial/Object/Goal/Long 四套件的算术平均。 各条 protocol 不同，不能合成一张总榜。
| 排名 | 模型 | 机构 | 参数量 | 开源权重 | Spatial | Object | Goal | Long | 平均 | 来源出处 |
|---|---|---|---|---|---|---|---|---|---|---|
| #1 | OpenVLA-OFT | Stanford / UC Berkeley | 7B | ✓ 开源 | 97.6% | 98.4% | 97.9% | 94.5% | 97.1% | OpenVLA-OFT Table I（腕部相机 + 本体感觉，过滤失败演示） ↗ third-person + wrist + proprio, modified dataset, one policy per suite, 500 trials/suite |
| #2 | OpenVLA-OFT (image + language) | Stanford / UC Berkeley | 7B | ✓ 开源 | 96.2% | 98.3% | 96.2% | 90.7% | 95.3% | OpenVLA-OFT Table I（仅第三人称图像 + 语言） ↗ third-person image + language, modified dataset, no wrist camera |
| #3 | UniVLA | OpenDriveLab | — | ✓ 开源 | 96.5% | 96.8% | 95.6% | 92% | 95.2% | UniVLA LIBERO table ↗ fine-tuned per suite; not the OpenVLA-OFT input setting |
| #4 | π0 | Physical Intelligence | 3.3B | ✓ 开源 | 96.8% | 98.8% | 95.8% | 85.2% | 94.2% | OpenVLA-OFT Table I 中的 π0 行（该表注明引自 π0 原文） ↗ third-person + wrist + proprio, modified dataset; number reprinted from the π0 paper |
| #5 | π0-FAST | Physical Intelligence | 3.3B | ✓ 开源 | 96.4% | 96.8% | 88.6% | 60.2% | 85.5% | OpenVLA-OFT Table I 中的 π0-FAST 行 ↗ third-person + wrist + proprio, modified dataset |
| #6 | SpatialVLA | Shanghai AI Laboratory | 4B | ✓ 开源 | 88.2% | 89.9% | 78.6% | 55.5% | 78.1% | SpatialVLA Table III ↗ fine-tuned per suite, 500 trials x 3 seeds; image-language setting comparable to OpenVLA 76.5, not to OFT 97.1 |
| #7 | OpenVLA | Stanford / UC Berkeley / TRI | 7B | ✓ 开源 | 84.7% | 88.4% | 79.2% | 53.7% | 76.5% | OpenVLA LIBERO table（OFT Table I 复引同一行） ↗ third-person image + language, modified dataset, original discrete-token fine-tune |
| #8 | Octo-Base | UC Berkeley RAIL | 93M | ✓ 开源 | 78.9% | 85.7% | 84.6% | 51.1% | 75.1% | OpenVLA LIBERO table 中的 Octo fine-tuned 行 ↗ fine-tuned; numbers reported by the OpenVLA paper, not the original Octo paper |
| #9 | Diffusion Policy | Columbia / Toyota Research Institute | — | ✓ 开源 | 78.3% | 92.5% | 68.3% | 50.5% | 72.4% | OpenVLA LIBERO table 中的 Diffusion Policy from-scratch 行 ↗ trained from scratch on the modified LIBERO set; reported by the OpenVLA paper |