Primotion Releases Embodied Base Model Paimon-0, Tops Multiple Embodied Intelligence Rankings
成立3个月便反超GPT-6 Astra！深圳黑马首次亮相
Shenzhen startup Primotion launched its embodied base model Paimon-0 within three months of its establishment, achieving top rankings in several benchmarks including RoboColiseum, RoboCasa365, and VLA-Arena.
The release of Paimon-0 by Primotion highlights a novel approach to embodied intelligence, combining model-based reinforcement learning with hardware-software integration. The model's performance across multiple benchmarks suggests strong capabilities in perception, planning, and execution, though the practical deployment readiness and safety validation remain to be assessed.
Source: QbitAI — Robotics · 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
task_success_rate
0.6613
View original evidence
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 S5task_success_rate
0.73
View original evidence
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 S7data_collection_cost
500–1,000
View original evidence
单个任务靠后训练把成功率提上去，很多时候已经能满足客户要求，真正卡脖子的是数采，一个任务动辄要采500到1000条数据，实验室采完还得去客户现场接着采。
Open source S21Reported duration
>2,000 hours
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.
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
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
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
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
单个任务靠后训练把成功率提上去，很多时候已经能满足客户要求，真正卡脖子的是数采，一个任务动辄要采500到1000条数据，实验室采完还得去客户现场接着采。
Open source S21
Source：QbitAI — Robotics · qbitai.com