QiYuan PrimeSim World Model Release
Overview
PrimeSim, a world model simulator developed by Qihuan's Embodied Research Center, outperforms existing simulators on two major benchmarks, EWMBench and WorldSimProbe. It leverages a two-stage training approach to transform a pre-trained video generation model into an interactive simulator capable of long-term consistent simulation. The model demonstrates strong cross-robotic-body transfer capabilities and achieves high performance in simulation fidelity and inference speed.
From Machine Heart — Robotics on WeChat
Event evidence and corrections
1 attributed source owners. Ownership does not establish independent confirmation. Quantities are reported separately and are never added together.
Environment: simulation
“今日，启元具身研究中心正式发布 PrimeSim，一款面向机器人操纵任务的世界模型仿真器。”
Exact source · revision 1jiqizhixin
Reported quantity · training_hours: 30110 hours · Basis not reported
“第一阶段：中训练｜从视频生成模型到机器人世界模型 模型训练基于 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，混合真实采集数据与传统仿真数据；其中人形机器人样本占比 54.8%，真实世界数据占比 84%。”
Exact source · revision 1jiqizhixin
Reported quantity · robot_platforms: 22 other · Basis not reported
“第一阶段：中训练｜从视频生成模型到机器人世界模型 模型训练基于 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，混合真实采集数据与传统仿真数据；其中人形机器人样本占比 54.8%，真实世界数据占比 84%。”
Exact source · revision 1jiqizhixin
Reported quantity · fps_with_cd: 10.04 other · Basis not reported
“在 RoboTwin 2.0 留出集、单张 A100 环境下，50 步去噪的因果自回归扩散基线速度仅 1.12 FPS，PSNR 26.15；CD 将去噪步数压缩至 4 步，速度提升至 10.04 FPS，PSNR 回落至 24.40；叠加 DMD 后，保持 4 步去噪，PSNR 回升至 24.87，LPIPS 由 0.193 优化至 0.176，相较 50 步教师模型整体提速约 9 倍。”
Exact source · revision 1jiqizhixin
Reported quantity · psnr_with_cd: 24.4 other · Basis not reported
“在 RoboTwin 2.0 留出集、单张 A100 环境下，50 步去噪的因果自回归扩散基线速度仅 1.12 FPS，PSNR 26.15；CD 将去噪步数压缩至 4 步，速度提升至 10.04 FPS，PSNR 回落至 24.40；叠加 DMD 后，保持 4 步去噪，PSNR 回升至 24.87，LPIPS 由 0.193 优化至 0.176，相较 50 步教师模型整体提速约 9 倍。”
Exact source · revision 1jiqizhixin
Reported quantity · psnr_with_dmd: 24.87 other · Basis not reported
“在 RoboTwin 2.0 留出集、单张 A100 环境下，50 步去噪的因果自回归扩散基线速度仅 1.12 FPS，PSNR 26.15；CD 将去噪步数压缩至 4 步，速度提升至 10.04 FPS，PSNR 回落至 24.40；叠加 DMD 后，保持 4 步去噪，PSNR 回升至 24.87，LPIPS 由 0.193 优化至 0.176，相较 50 步教师模型整体提速约 9 倍。”
Exact source · revision 1jiqizhixin
Reported quantity · lpips_with_dmd: 0.176 other · Basis not reported
“在 RoboTwin 2.0 留出集、单张 A100 环境下，50 步去噪的因果自回归扩散基线速度仅 1.12 FPS，PSNR 26.15；CD 将去噪步数压缩至 4 步，速度提升至 10.04 FPS，PSNR 回落至 24.40；叠加 DMD 后，保持 4 步去噪，PSNR 回升至 24.87，LPIPS 由 0.193 优化至 0.176，相较 50 步教师模型整体提速约 9 倍。”
Exact source · revision 1jiqizhixin
Report timeline
Follow attributed reports and material updates.
- Machine Heart — Robotics on WeChatSignalPrimeSim: A World Model Simulator for Robotics That Outperforms Existing Benchmarks
PrimeSim, a world model simulator developed by Qihuan's Embodied Research Center, outperforms existing simulators on two major benchmarks, EWMBench and WorldSimProbe. It leverages a two-stage training approach to transform a pre-trained video generation model into an interactive simulator capable of long-term consistent simulation. The model demonstrates strong cross-robotic-body transfer capabilities and achieves high performance in simulation fidelity and inference speed.
Event coverage history
There is not enough continuous observation data to show a trend.
Timezone · UTC
Article dates follow your selected timezone. Briefing editions use Hong Kong time (UTC+8).