NVIDIA releases the VFlash visual planning decoding framework, accelerating robot action decoding by up to 12x
NVIDIA has released VFlash, a general-purpose decoding framework for fast visual planning that uses parallel decoding of complete visual primitives to replace token-by-token coordinate generation, enabling prediction of robot grasp positions and motion trajectories in a single pass. Under H100 batch size 1, single-pass generation speed improves by about 25x; across four MolmoSpaces tasks, the world action model achieves an average success rate of 70.2% versus a 62.6% baseline. Jim Fan shared that it achieves up to 12x action decoding acceleration, 5.1x overall robot inference acceleration, and up to 61% reduction in 3D trajectory error.
Highlights
Zhongqing and Peking University release PyRUA-REAL: real-machine self-evolution reaches 84.42% on RoboDojo
Zhongqing Robotics and a Peking University research team jointly released the paper PyRUA-REAL. With foundation model parameters frozen, two coding agents—a researcher and an engineer—repeatedly explore, diagnose, and modify code, distilling effective practices into skills and memory.
Northwestern, CMU, and others propose WOVEN: using 2k samples to improve 22 benchmarks across domains and providing a training recipe
A team from Northwestern, CMU, UNC Chapel Hill, and Amazon, in the paper "WOVEN: Weaving Visual World Modeling into Multimodal LLMs," defines world modeling as estimation of the state transition p(s′|s,a) and builds the WOVEN training and evaluation base of 36K questions.
机器之心 · 微信公众号Source ↗Yuance Future releases Anchor 1, a whole-body control model for humanoid robots
Yuance Future has released Anchor 1, a whole-body control model positioned as the System 0 cerebellum layer within a whole-body intelligence architecture. It uses five anchor points—the pelvis, left and right wrists, and left and right ankles—as commands and does not rely on online retargeting.
机器之心 · 微信公众号Source ↗
Model release/update
NVIDIA releases the VFlash visual planning decoding framework, accelerating robot action decoding by up to 12x
NVIDIA has released VFlash, a general-purpose decoding framework for fast visual planning that uses parallel decoding of complete visual primitives to replace token-by-token coordinate generation, enabling prediction of robot grasp positions and motion trajectories in a single pass. Under H100 batch size 1, single-pass generation speed improves by about 25x; across four MolmoSpaces tasks, the world action model achieves an average success rate of 70.2% versus a 62.6% baseline. Jim Fan shared that it achieves up to 12x action decoding acceleration, 5.1x overall robot inference acceleration, and up to 61% reduction in 3D trajectory error.
机器人前瞻 · 微信公众号Source ↗Yuance Future releases Anchor 1, a whole-body control model for humanoid robots
Yuance Future has released Anchor 1, a whole-body control model positioned as the System 0 cerebellum layer within a whole-body intelligence architecture. It uses five anchor points—the pelvis, left and right wrists, and left and right ankles—as commands and does not rely on online retargeting.
机器之心 · 微信公众号Source ↗
Product release/update
Shanghai AI Laboratory launches the Sage·LabOS laboratory operating system
Shanghai AI Laboratory has launched the laboratory operating system Sage·LabOS, which uses ShuSheng SCP-Lab as a unified interface and the ShuSheng physical world model W0 to drive robots in completing experimental operations, connecting heterogeneous devices for autonomous execution and data return. The system has four core capabilities: capability discovery, task orchestration, execution scheduling, and data feedback. It has completed five rounds of dry-wet closed-loop iterations in a formulation optimization task at Lingang Laboratory, automatically synthesizing and validating 335 formulations and raising the encapsulation rate from 67% to 98%.
Industry news
Dream Labs launches Dream API, with the robot brain billed at $3/hour
Dream Labs, spun out of NVIDIA's GR00T team, has completed a $90 million seed round and launched Dream API, a publicly registrable, explicitly priced robot foundation model API service.
机器人产业应用 · 微信公众号Source ↗Zhijian Dynamics and Hangcha Group deploy a robot sorting line at a supermarket front warehouse
On November 3, Zhijian Dynamics and Hangcha Group completed the deployment of the first robot sorting line at a supermarket front warehouse, connecting the full process from AMR merged feeding, robot autonomous planning and dexterous grasping and turning, code scanning verification, to order distribution. The line uses the iX robot body customized by Zhijian Dynamics for Hangcha and, relying on visual perception and VLA task planning, handles boxed, bagged, bottled, and other forms of goods without the need to model each SKU separately. Measured single-machine sorting efficiency reaches up to 450 items/hour, with accuracy stable above 98%, and it can support 24-hour continuous fault-free operation at full load.
至简动力 · 微信公众号Source ↗Six AgiBot G2 robots complete a 24-hour pilot at a Shanghai convenience store, performing 1,730 tasks in total
Six AgiBot G2 wheeled humanoid robots completed a 24-hour pilot at an operating convenience store in Xuhui District, Shanghai, performing 1,730 tasks in total, with four of them passing the final assessment.
Tashi Intelligent Navigation signs with Aptiv to deploy hundred-unit-scale A-series industrial embodied robots by year-end
Tashi Intelligent Navigation and Aptiv signed an agreement during the 9th China International Import Expo to deepen strategic cooperation in embodied intelligence. Aptiv will complete the deployment of hundred-unit-scale Tashi A-series industrial embodied intelligent robots by the end of this year, covering 8 major production processes and 33 application scenarios. The first batch of robots has already been put into use on real production lines at Aptiv's smart factory. Products such as Aptiv's high-performance computing platform will be the first to serve as the computing base for Tashi's embodied robots, with mass production planned within 2026. The two parties will also continue to accumulate data and validate applications around real industrial scenarios based on Tashi's general embodied large model AWE.
它石智航 · 微信公众号Source ↗
Paper research
Zhongqing and Peking University release PyRUA-REAL: real-machine self-evolution reaches 84.42% on RoboDojo
Zhongqing Robotics and a Peking University research team jointly released the paper PyRUA-REAL. With foundation model parameters frozen, two coding agents—a researcher and an engineer—repeatedly explore, diagnose, and modify code, distilling effective practices into skills and memory.
Northwestern, CMU, and others propose WOVEN: using 2k samples to improve 22 benchmarks across domains and providing a training recipe
A team from Northwestern, CMU, UNC Chapel Hill, and Amazon, in the paper "WOVEN: Weaving Visual World Modeling into Multimodal LLMs," defines world modeling as estimation of the state transition p(s′|s,a) and builds the WOVEN training and evaluation base of 36K questions.
机器之心 · 微信公众号Source ↗