Skip to content
QbitAI — Robotics on WeChat· 关注前沿科技·· 7 hours agoSignalEditorial score92

Paimon-0 Leads in Multiple Embodied Benchmark Tests, Shenzhen Startup Primotion Makes Debut

成立3个月便反超GPT-6 Astra！深圳黑马首次亮相

Summary

Primotion, a Shenzhen-based company established just three months ago, has launched its embodied base model Paimon-0, which has achieved first place in several benchmarks including RoboColiseum, RoboCasa365, and VLA-Arena. The model employs a World Action Model (WAM) and Inverse Dynamics Model (IDM) architecture, combined with FPGA hardware control to achieve low latency. The company enhances the model's generalization capabilities through data closed-loop and self-improvement systems (RSI), and plans to open-source the Progress Model Paimon-Eval.

Source: QbitAI — Robotics on WeChat · Read original article ↗

English translation unavailable. Choose the original article to read it.

What the source reports

Publisher-reported claims, with original evidence. These results have not been independently verified by RoboSignal.

Reported numbers

  • tasks

    50

    View original evidence
    RoboCasa365 是面向通用机器人策略的大规模多任务基准，包含2500+仿真厨房场景、365项日常任务、2000+小时交互数据，选取50项任务，从原子可见、组合可见、组合不可见三层检验泛化能力，Paimon-0以58.1的综合分数登顶。
    Open source S6

  • total_score

    46.599

    View original evidence
    同时Paimon-0参加启元机器人主办的 IROS Workshop 挑战赛，赛道聚焦双臂机器人策略，包含仿真线上评测与真机线下验证，炎曦动力最终以46.599总分夺冠，验证了端到端策略在双臂操作上的能力。
    Open source S8

  • total_score

    58.1

    View original evidence
    RoboCasa365 是面向通用机器人策略的大规模多任务基准，包含2500+仿真厨房场景、365项日常任务、2000+小时交互数据，选取50项任务，从原子可见、组合可见、组合不可见三层检验泛化能力，Paimon-0以58.1的综合分数登顶。
    Open source S6
  • paper: not_reported
  • code: not_reported
  • weights: not_reported
  • dataset: not_reported
Source excerpts and review record

Automatically extracted; no manual editorial approval recorded.

RoboCasa365 是面向通用机器人策略的大规模多任务基准，包含2500+仿真厨房场景、365项日常任务、2000+小时交互数据，选取50项任务，从原子可见、组合可见、组合不可见三层检验泛化能力，Paimon-0以58.1的综合分数登顶。

Open source S6

同时Paimon-0参加启元机器人主办的 IROS Workshop 挑战赛，赛道聚焦双臂机器人策略，包含仿真线上评测与真机线下验证，炎曦动力最终以46.599总分夺冠，验证了端到端策略在双臂操作上的能力。

Open source S8

VLA-Arena 为通用具身模型开源评测榜单，围绕安全性、抗干扰能力做系统化测评，支持跨模型对比，Paimon-0取得0.73平均成功率，排名第一。

Open source S7

要知道，这个榜最近可是刚把GPT-6 Astra、Claude Opus 5.5、Gemini 4 Argon这些通用多模态大模型请进场同台比试的： 虽说大模型在指令跟随和空间理解上分数确实亮眼，可一到要真刀真枪动手的通用操作子榜，Paimon-0的0.6613依然是全场最高，扰动适应子榜也由它领跑。

Open source S5

然后这个系列里的首款模型 Paimon-0 初次亮相，就直接拿下了四项第一： 先看含金量最高的 RoboColiseum ，这个由 Hugging Face、NVIDIA、清华大学等全球近 20 家高校、企业、科研机构共建的榜单，最近在具身圈热度居高不下；Paimon-0 以 0.7754 分排在第一。

Open source S4

Source：QbitAI — Robotics on WeChat · mp.weixin.qq.com

#Model release#Robot policies#Imitation learning#Reinforcement learning#Sim-to-real#Evaluation
Timezone · UTC

Article dates follow your selected timezone. Briefing editions use Hong Kong time (UTC+8).