Chinese regulators block SoftBank's $5.4 billion acquisition of ABB's robotics business
This week embodied intelligence models were updated intensively; NVIDIA VFlash replaces token-by-token generation with parallel decoding, speeding up robot action decoding by up to 12 times. On the industry side, Chinese regulators blocked SoftBank's $5.375 billion acquisition of ABB's robotics business, and Walmart and Symbotic expanded their agreement with a commitment of about $6 billion in initial investment. On the research side, the focus is on data and simulation, with PrimeSim, HALO, and others using large-scale demonstrations and synthetic data to advance general world models and zero-shot real-robot transfer.
This week embodied intelligence models were updated intensively; NVIDIA VFlash replaces token-by-token generation with parallel decoding, speeding up robot action decoding by up to 12 times. On the industry side, Chinese regulators blocked SoftBank's $5.375 billion acquisition of ABB's robotics business, and Walmart and Symbotic expanded their agreement with a commitment of about $6 billion in initial investment. On the research side, the focus is on data and simulation, with PrimeSim, HALO, and others using large-scale demonstrations and synthetic data to advance general world models and zero-shot real-robot transfer.
Highlights
PrimeSim: a general-purpose world model simulator based on large-scale cross-embodiment demonstrations
The research team proposed PrimeSim, an action-conditioned multi-view world model simulator, mid-trained on robot demonstration data covering 22 platforms and 30,110 hours, and released model weights, inference code, and training code.
HuggingFace Daily PapersSource ↗Zhongqing and Peking University release PyRUA-REAL: real-robot self-evolution reaches 84.42% on RoboDojo
Zhongqing Robotics and a Peking University research team jointly released the paper PyRUA-REAL, in which, with base model parameters frozen, two coding agents—a researcher and an engineer—repeatedly explore, diagnose, and modify code, distilling effective practices into skills and memory.
众擎机器人 · 微信公众号Source ↗The first public accounts of humanoid robots: an in-depth analysis on the eve of Agility Robotics' listing
Agility Robotics will list on Nasdaq via SPAC on November 5 under the ticker AGLT, with a pre-money valuation of $2.5 billion and confirmed revenue of only $1.782 million in 2025, about 63.5% of which came from related parties.
钛媒体Source ↗
Model Releases/Updates
This week, multiple institutions launched world action models and whole-body control models, covering navigation, mobile manipulation, and dexterous manipulation, with an emphasis on open source and public technical reports.
NVIDIA releases VFlash visual planning decoding framework, accelerating robot action decoding by up to 12x
NVIDIA released VFlash, a general-purpose decoding framework for fast visual planning, which uses parallel decoding of complete visual primitives to replace token-by-token coordinate generation, enabling prediction of robot grasp positions and motion trajectories in one pass. At batch size 1 on H100, single-pass generation speed improves by about 25x, and across four MolmoSpaces tasks the world action model achieves an average success rate of 70.2% versus a baseline of 62.6%. 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 ↗LiberAI releases world action model Liber-0, and publishes technical report and online simulator
LiberAI, founded by Liu Songming, released the next-generation world action model Liber-0, which uses three expert streams for language, vision, and action together with Multimodal Attention to jointly predict future frames and actions.
机器之心 · 微信公众号Source ↗Fysics AI and Fudan propose Fysiverse-UniWAM unified world action model, ranking first in both navigation and mobile manipulation across 24 real-robot tests
Fysics AI, together with Fudan University CITLab, proposed the unified world action model Fysiverse-UniWAM, using a single model to connect autonomous navigation, mobile manipulation, and fixed-workstation dexterous manipulation.
机器之心 · 微信公众号Source ↗Yuance Future releases humanoid whole-body control model Anchor 1
Yuance Future released the whole-body control model Anchor 1, positioned as the System 0 cerebellum layer in a whole-body intelligence architecture, using five anchor points—pelvis, left and right wrists, and left and right ankles—as commands without relying on online retargeting.
机器之心 · 微信公众号Source ↗Ant Lingbo embodied brain full stack 2.0 debuts at Wuzhen Summit, completes autonomous prescription retrieval in narrow pharmacy aisle
Ant Lingbo'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, drug identification, and sorting and delivery within an 80-centimeter-wide pharmacy aisle; the same solution has been operating routinely in multiple Shanghai Guoda Pharmacy stores.
AI前线 · 微信公众号Source ↗
Industry News
Frequent capital and supply chain moves: Agility Robotics is about to list via SPAC, and Leaderdrive's humanoid reducer orders increase to 400,000 units.
Chinese regulators block SoftBank's $5.4 billion acquisition of ABB's robotics business
China's Foreign Investment Security Review Working Mechanism Office prohibited SoftBank's $5.375 billion acquisition of ABB's robotics business, citing reasons involving robot data being transferred overseas and used to train AI models. ABB Robotics' orders from China in fiscal 2025 were $928,000, up from $662,000 in 2024, and its China business grew 32% in the third quarter of 2026, with about 90% of the robots produced at its Shanghai factory for local customers manufactured locally. ABB said it remains in close communication with SoftBank and is advancing the transaction.
The Robot ReportSource ↗The first public accounts of humanoid robots: an in-depth analysis on the eve of Agility Robotics' listing
Agility Robotics will list on Nasdaq via SPAC on November 5 under the ticker AGLT, with a pre-money valuation of $2.5 billion and confirmed revenue of only $1.782 million in 2025, about 63.5% of which came from related parties.
钛媒体Source ↗Walmart and Symbotic expand commercial agreement to add automation capabilities to 25 U.S. regional distribution centers
Walmart and Symbotic announced an expanded commercial agreement that will bring more automation capabilities to 25 U.S. regional distribution centers and support more opportunities across Walmart's global supply chain network. The agreement includes an initial investment commitment of about $6 billion, covering multi-year work at multiple locations. Symbotic said its integrated technologies, including the BreakPack each-picking solution and the SymMicro micro-fulfillment platform, are supported by AI software for scheduling and optimizing the flow of millions of cases per day.
Robotics 24/7Source ↗Leaderdrive's first three quarters revenue up 45%, humanoid reducer orders reach 400,000 units
Leaderdrive's revenue in the first three quarters of 2026 was 591 million yuan, up 45.30% year on year, with net profit attributable to shareholders of 122 million yuan, up 30.33% year on year, and humanoid-related reducer orders increased from about 100,000 units the previous year to 400,000 units. The company's annual harmonic reducer capacity has reached 1 million units, with a year-end target of 1.5 million units, but first-half gross margin was 32.26%, down 2.51 percentage points year on year, as new production line commissioning and small-batch production of multiple models are weighing on profitability. The article also compares the catch-up pace of domestic manufacturers such as Zhongda Drede, Huandong Technology, and Laifu Harmonic.
人形大讲堂 · 微信公众号Source ↗
Paper Research
Research focused on large-scale datasets and simulation training, with tactile datasets, human video motion data, and closed-loop simulation real-world testing becoming this week's highlights.
PrimeSim: a general-purpose world model simulator based on large-scale cross-embodiment demonstrations
The research team proposed PrimeSim, an action-conditioned multi-view world model simulator, mid-trained on robot demonstration data covering 22 platforms and 30,110 hours, and released model weights, inference code, and training code.
HuggingFace Daily PapersSource ↗Zhongqing and Peking University release PyRUA-REAL: real-robot self-evolution reaches 84.42% on RoboDojo
Zhongqing Robotics and a Peking University research team jointly released the paper PyRUA-REAL, in which, with base model parameters frozen, two coding agents—a researcher and an engineer—repeatedly explore, diagnose, and modify code, distilling effective practices into skills and memory.
众擎机器人 · 微信公众号Source ↗HALO: a humanoid robot autonomy system trained on synthetic simulation data achieves zero-shot real-robot transfer
The research team proposed HALO, a humanoid robot autonomy system, which synthesizes vision-language-action data in about one thousand procedurally generated home scenes to eliminate reliance on teleoperation collection. HALO learns skills such as sitting on a chair, navigating to a specified object, turning sideways to pass through a half-open door, crouching and crawling under a table, and pushing a box, and can combine them to complete long-horizon tasks. Without having seen real home scenes during training, HALO transferred zero-shot to the physical world, achieving a 91% success rate across 360 trials in six unmapped real environments.
arXiv cs.ROSource ↗Movens-Zero: scaling humanoid motion foundation models with large-scale human video
The research team proposed Movens-Zero, a scalable understand-generate-execute architecture that supports real-time closed-loop humanoid motion generation under text, speech, music, and video inputs. Its accompanying dataset Movens-5M was built from internet videos through clip generation, quality filtering, semantic annotation, motion recovery, and physics refinement via humanoid retargeting, containing 15 million human motion sequences and 5 million physically executable humanoid motion sequences, totaling more than 10,000 hours.
arXiv cs.ROSource ↗TacClaw-Infinity: a large-scale vision-tactile-language-action dataset for dexterous manipulation
The research team released TacClaw-Infinity, a large-scale structured vision-tactile-language-action (VTLA) dataset for contact-intensive manipulation, with more than 10,000 hours, 1 million episodes, and 1 billion frames, covering more than 1,000 task types.
arXiv cs.ROSource ↗cuRoboV2-Fast: a GPU-accelerated planner for torque-limited collision-free motion
The authors proposed cuRoboV2-Fast, a GPU-accelerated planner for collision-free, payload-aware, and torque-limited motion, reducing iteration counts through B-spline trajectory optimization and using TimeOpt to optimize traversal speed along fixed paths under velocity, acceleration, jerk, and torque constraints.
arXiv cs.ROSource ↗Any3R open-sourced: one model uniformly handles 3D reconstruction and pose estimation for pinhole, fisheye, and panoramic cameras
South Dakota State University and other institutions proposed Any3R, which describes arbitrary central cameras with a dense ray field and can reconstruct a 3D point cloud and estimate pose from mixed-camera images in a single forward pass. The method replaces the row and column indices of rotary position encoding with angular coordinates and weights pixel loss by solid angle, enabling one set of weights to cover pinhole, over-180° fisheye, and 360° panoramic cameras. On benchmarks across 13 central camera types and 3 datasets, its reconstruction accuracy and pose accuracy both lead, and the inference code, 13 camera models, and weights have been open-sourced.
3D视觉工坊 · 微信公众号Source ↗Northwestern, CMU, and others propose WOVEN: using 2k samples to improve 22 benchmarks across domains and provide a training recipe
In the paper "WOVEN: Weaving Visual World Modeling into Multimodal LLMs," teams from Northwestern, CMU, UNC Chapel Hill, and Amazon define world modeling as the estimation of state transitions p(s′|s,a) and build the WOVEN training and evaluation base of 36K questions.
机器之心 · 微信公众号Source ↗Dong Hao's team at Peking University proposes One Scan is Enough: completing post-training for embodied navigation with a single scan
Dong Hao's team at Peking University publishes the paper "One Scan is Enough," which uses a LiDAR-equipped iPhone to scan a target environment once, automatically generates navigation tasks, and uses reconstructed scene geometry for reinforcement post-training, without any manual teleoperation demonstrations.
甲子光年 · 微信公众号Source ↗RoboTryOn: Synthesizing dexterous robot demonstrations from human videos via virtual try-on
RoboTryOn proposes treating human-to-robot transfer as virtual try-on: preserving the demonstrated motion, replacing the human with a 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) with fused depth features to infer action labels from the synthesized video for policy learning.
HuggingFace Daily PapersSource ↗GE-Sim 2.0 and π0.5 closed-loop simulation tests: why reward hacking emerges in VLA training
The Xbotics Embodied Intelligence Lab builds a closed-loop simulation pipeline using the Zhiyuan GE-Sim 2.0 world model and the OpenPI π0.5 policy, and measures that the world model generates about 4.33 frames per second, the pure-policy closed loop runs at about 3.20 frames per second, and after adding MiMo online scoring it drops to about 0.88 frames per second.
Xbotics具身智能实验室 · 微信公众号Source ↗