PrimeSim: A general-purpose world model simulator based on large-scale cross-embodiment demonstrations
The research team proposes PrimeSim, an action-conditioned multi-view world model simulator, mid-trains it on robot demonstration data covering 22 platforms and 30,110 hours, and releases the model weights, inference code, and training code.
Highlights
Ant Group's Lingbo embodied brain full-stack 2.0 debuts at the Wuzhen Summit, completing autonomous medicine retrieval in narrow pharmacy aisles
Ant Group's open-source embodied brain full-stack 2.0 was exhibited at the 2026 World Internet Conference Wuzhen Summit, where a robot completed order acceptance, medicine identification, and sorting and delivery within an 80-centimeter-wide pharmacy aisle; the same solution is already running routinely in multiple Shanghai Guoda Pharmacy stores.
AI前线 · 微信公众号Source ↗RoboTryOn: Synthesizing human videos into dexterous robot demonstrations via virtual try-on
RoboTryOn treats human-to-robot conversion as virtual try-on, preserving the demonstrated motion, replacing the human with the target robot specified by a reference image, and keeping the task, objects, and background unchanged. The method autoregressively converts keyframes to guide video synthesis, conditions motion on the human hand skeleton, and then uses an inverse dynamics model (IDM) fused with depth features to infer action labels from the synthesized videos for policy learning.
HuggingFace Daily PapersSource ↗D-Robotics and the University of Tokyo release RawVLA: freezing the VLA and training only the imaging front end raises success rate on low-light dual-arm tasks from 0 to 66.5%
D-Robotics, together with the University of Tokyo and other institutions, released the RawVLA technical approach, which adds a learnable neural ISP module before the VLA policy, keeps the VLA model frozen, and trains the imaging front end in reverse using action error. Real-robot tests with π0.5 on four dual-arm tasks raised the average success rate under low light from 0% with the default ISP to 66.5%, versus 75.0% under normal light, where the default ISP achieved 76.5%.
未来机器人网 · 微信公众号Source ↗
Model releases/updates
Ant Group's Lingbo embodied brain full-stack 2.0 debuts at the Wuzhen Summit, completing autonomous medicine retrieval in narrow pharmacy aisles
Ant Group's open-source embodied brain full-stack 2.0 was exhibited at the 2026 World Internet Conference Wuzhen Summit, where a robot completed order acceptance, medicine identification, and sorting and delivery within an 80-centimeter-wide pharmacy aisle; the same solution is already running routinely in multiple Shanghai Guoda Pharmacy stores.
AI前线 · 微信公众号Source ↗Future Not Far releases Self-Evolving WAM self-evolving world action model
Future Not Far releases the Self-Evolving WAM self-evolving world action model, integrating joint world-action modeling, interactive experience learning, and training-effect feedback into a single system. Across update attempts from round 0 to round 30, the autonomous full-task success rates of four household tasks all improved: clothes folding rose from 54.6% to 96.8%, clothes drying rose from 46.2% to 91.2%, making pasta rose from 42.7% to 89.5%, and stir-fried pork with chili rose from 38.2% to 68.9%.
未来不远 · 微信公众号Source ↗
Product releases/updates
Origin Force Machine upgrades the Dexmal MaaS platform, connecting the chain from embodied model training to real-robot deployment
Origin Force Machine has comprehensively upgraded Dexmal MaaS, connecting the complete chain from data collection and policy training to cloud inference and real-robot deployment; the platform is based on its self-developed DM0.5 embodied foundation model. The platform provides three training pipelines—SFT, DAgger, and RECAP—corresponding respectively to building foundational capabilities, adapting to real environments, and optimizing failure cases, and supports DexOS Lite and MaaS API access for custom robots.
Dexmal 原力灵机 · 微信公众号Source ↗UBTech releases Thinker-Sim embodied intelligence simulation platform
UBTech releases Thinker-Sim, an embodied intelligence simulation platform centered on Real2Sim2Real, covering three parts: scene reconstruction and asset generation, simulation data synthesis, and robot capability evaluation. The platform claims that its self-developed rendering engine improves rendering efficiency by 67.3%, synthesizes 390 hours of simulation data per day per node (8 GPUs), and improves SIL/HIL communication efficiency by more than 400%. The platform also provides standardized task templates, visual workflows, and natural-language task orchestration based on Ulab-Agent.
Industry news
Agility Robotics completes merger with Churchill Capital Corp XI and will trade on Nasdaq under AGLT
Agility Robotics has completed its previously announced business combination with Churchill Capital Corp XI; the combined company's common stock and warrants are expected to begin trading on Nasdaq on November 5, 2026, under AGLT and AGLTW, respectively.
Paper research
PrimeSim: A general-purpose world model simulator based on large-scale cross-embodiment demonstrations
The research team proposes PrimeSim, an action-conditioned multi-view world model simulator, mid-trains it on robot demonstration data covering 22 platforms and 30,110 hours, and releases the model weights, inference code, and training code.
HuggingFace Daily PapersSource ↗RoboTryOn: Synthesizing human videos into dexterous robot demonstrations via virtual try-on
RoboTryOn treats human-to-robot conversion as virtual try-on, preserving the demonstrated motion, replacing the human with the target robot specified by a reference image, and keeping the task, objects, and background unchanged. The method autoregressively converts keyframes to guide video synthesis, conditions motion on the human hand skeleton, and then uses an inverse dynamics model (IDM) fused with depth features to infer action labels from the synthesized videos for policy learning.
HuggingFace Daily PapersSource ↗D-Robotics and the University of Tokyo release RawVLA: freezing the VLA and training only the imaging front end raises success rate on low-light dual-arm tasks from 0 to 66.5%
D-Robotics, together with the University of Tokyo and other institutions, released the RawVLA technical approach, which adds a learnable neural ISP module before the VLA policy, keeps the VLA model frozen, and trains the imaging front end in reverse using action error. Real-robot tests with π0.5 on four dual-arm tasks raised the average success rate under low light from 0% with the default ISP to 66.5%, versus 75.0% under normal light, where the default ISP achieved 76.5%.
未来机器人网 · 微信公众号Source ↗