China's regulator blocks SoftBank's $5.4 billion acquisition of ABB's robotics business
China's Foreign Investment Security Review 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; about 90% of the robots produced at its Shanghai factory for local customers are locally manufactured. ABB said it remains in close communication with SoftBank and is advancing the transaction.
Highlights
HALO: a humanoid robot autonomy system trained on simulated synthetic data achieves zero-shot real-world transfer
The research team proposes the humanoid robot autonomy system HALO, which synthesizes vision-language-action data in about one thousand procedurally generated home scenes, eliminating reliance on teleoperation collection. HALO learned skills such as sitting on a chair, navigating to a specified object, sidestepping through a half-open door, crouching and crawling under a table, and pushing a box, and can combine them to complete long-horizon tasks. It never saw real home scenes during training, yet HALO achieved a 91% success rate in 360 trials across six unmapped real environments through zero-shot transfer to the physical world.
arXiv cs.ROSource ↗LiberAI releases the world action model Liber-0 and publishes a technical report and online simulator
LiberAI, founded by Liu Songming, released the new-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 ↗cuRoboV2-Fast: a GPU-accelerated planner for torque-limited collision-free motion
The authors propose the GPU-accelerated planner cuRoboV2-Fast for collision-free, payload-aware, and torque-limited motion, reducing iteration counts through B-spline trajectory optimization and using TimeOpt to optimize the traversal speed of a fixed path under velocity, acceleration, jerk, and torque constraints.
arXiv cs.ROSource ↗
Model releases/updates
LiberAI releases the world action model Liber-0 and publishes a technical report and online simulator
LiberAI, founded by Liu Songming, released the new-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 ↗
Product releases/updates
Atomrobot releases the bionic five-finger dexterous hand Tianqiao No. 1
Atomrobot launched its self-developed bionic five-finger dexterous hand Tianqiao No. 1, featuring tendon-based flexible transmission and an inverted drive unit, with an overall hand height of 221 mm, a single finger length of 126 mm, 7 degrees of freedom, and retained thumb opposition capability. The hand comes standard with five fingertip sensors and one palm sensor, and with a self-developed tendon nonlinear compensation algorithm achieves ±0.1 mm repeat positioning accuracy. The product also ships with a native motion control framework supporting switching between position control and compliant force control, and prototype evaluation and business inquiries are open starting today.
Industry developments
China's regulator blocks SoftBank's $5.4 billion acquisition of ABB's robotics business
China's Foreign Investment Security Review 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; about 90% of the robots produced at its Shanghai factory for local customers are locally manufactured. ABB said it remains in close communication with SoftBank and is advancing the transaction.
Paper research
HALO: a humanoid robot autonomy system trained on simulated synthetic data achieves zero-shot real-world transfer
The research team proposes the humanoid robot autonomy system HALO, which synthesizes vision-language-action data in about one thousand procedurally generated home scenes, eliminating reliance on teleoperation collection. HALO learned skills such as sitting on a chair, navigating to a specified object, sidestepping through a half-open door, crouching and crawling under a table, and pushing a box, and can combine them to complete long-horizon tasks. It never saw real home scenes during training, yet HALO achieved a 91% success rate in 360 trials across six unmapped real environments through zero-shot transfer to the physical world.
arXiv cs.ROSource ↗cuRoboV2-Fast: a GPU-accelerated planner for torque-limited collision-free motion
The authors propose the GPU-accelerated planner cuRoboV2-Fast for collision-free, payload-aware, and torque-limited motion, reducing iteration counts through B-spline trajectory optimization and using TimeOpt to optimize the traversal speed of a fixed path under velocity, acceleration, jerk, and torque constraints.
arXiv cs.ROSource ↗PyRUA-DataFarm: using LLM agents instead of teleoperation to generate real usable robot data
PyRUA-DataFarm proposes having coding agents reconstruct tasks in simulation and video models render them into real usable data, replacing teleoperation collection of robot demonstrations. On five real robot tasks, policies trained only on this data achieved 93/100 successes, compared with 84/100 for policies trained on teleoperation data; under unseen tablecloths and colored lighting, the former maintained 87 successes while the latter dropped to 46.
HuggingFace Daily PapersSource ↗
Brief updates
NVIDIA proposes a Cosmos-based interactive world simulation framework
NVIDIA proposes a shared framework that adapts the Cosmos world foundation model into real-time interactive world simulation for Physical AI. The framework can stream visual observations based on action and scene control flows, supporting both predicting the visual consequences of actions and generating observations from physical states maintained by an external simulator. The authors demonstrated it in camera navigation, robotic manipulation, controllable video transfer, and multi-view autonomous driving simulation, and released the code and model checkpoints.
arXiv cs.ROSource ↗AutoDex: a system for automatically collecting real-world dexterous grasping data
AutoDex is a system that collects real-world dexterous grasping data without continuous human operation, automatically planning and executing grasps, labeling results, and resetting objects for continuous collection. The research team used Allegro and Inspire hands to collect more than 3,500 grasp trials on 100 objects, recording successes and failures along with synchronized visual observations and robot states; collecting 500 trajectories required about 0.2 human-hours, compared with 9.3 human-hours for teleoperation.
arXiv cs.ROSource ↗