QiYuan Embodied Research Center Releases PrimeSim World Model Simulator
启元机器人发布世界模型仿真器
On November 4, QiYuan Embodied Research Center released PrimeSim, a world model simulator based on pre-trained video generation models modified through 'in-training' and 'post-training' stages. The system supports robot manipulation strategy training by maintaining geometric and semantic consistency across multiple camera views while adhering to action commands. Weights, inference code, and training code are scheduled for open-source release on November 6. In EWMBench benchmarking, PrimeSim achieved a top score of 6.09. Trained on 30,110 hours of robot demonstration data covering 22 platforms, it demonstrated strong cross-embodiment adaptation with minimal gradient updates for new robots.
Source: Chinese robotics — Research news discovery · 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
training_hours
30,110 hours
View original evidence
“中训练”阶段，模型基于30110小时精选机器人演示数据，覆盖人形、双臂、单臂、手持UMI设备等22种异构机器人平台，其中人形机器人样本占比54.8%、真实世界数据占比84%，并围绕统一动作表征、帧对齐双路径条件注入、视角感知交错注意力三项核心设计实现跨本体兼容与控制信号强绑定。
Open source S12xArm7_gradient_steps
200
View original evidence
在少样本本体迁移方面，“中训练”赋予模型较强的跨机器人本体适配能力：适配xArm7、ViperX300、UMI夹爪分别仅需200、100、200步梯度更新。
Open source S18ViperX300_gradient_steps
100
View original evidence
在少样本本体迁移方面，“中训练”赋予模型较强的跨机器人本体适配能力：适配xArm7、ViperX300、UMI夹爪分别仅需200、100、200步梯度更新。
Open source S18UMI_claw_gradient_steps
200
View original evidence
在少样本本体迁移方面，“中训练”赋予模型较强的跨机器人本体适配能力：适配xArm7、ViperX300、UMI夹爪分别仅需200、100、200步梯度更新。
Open source S18AgileXPiper_finetune_steps
≈200
- paper: not_reported
- code: Reported pending Open source S4
- weights: Reported pending Open source S4
- dataset: not_reported
- code: not_reported
Source excerpts and review record
Automatically extracted; no manual editorial approval recorded.
“中训练”阶段，模型基于30110小时精选机器人演示数据，覆盖人形、双臂、单臂、手持UMI设备等22种异构机器人平台，其中人形机器人样本占比54.8%、真实世界数据占比84%，并围绕统一动作表征、帧对齐双路径条件注入、视角感知交错注意力三项核心设计实现跨本体兼容与控制信号强绑定。
Open source S12
在少样本本体迁移方面，“中训练”赋予模型较强的跨机器人本体适配能力：适配xArm7、ViperX300、UMI夹爪分别仅需200、100、200步梯度更新。
Open source S18
在未见于训练集的AgileXPiper场景中，经单卡A100约200步微调，布料形变即可跟随动作指令，且多视角状态保持一致。
Open source S19
相关权重、推理代码与训练代码将于11月6日开源。
Open source S4
Source：Chinese robotics — Research news discovery · app.xinhuanet.com