China blocked SoftBank's acquisition of ABB Robotics due to concerns that training data for AI models could be transferred overseas. This action stems from regulatory worries regarding the transfer of robotics data, which is now viewed as strategically important for physical AI development. Other countries are also taking action to protect sovereign physical AI, with examples including U.S. bans on certain autonomous robots. The issue highlights growing concerns about robotics data, intellectual property, and local manufacturing capabilities in cross-border mergers. Ultimately, the situation reflects broader geopolitical tensions concerning the convergence of automation, data, and artificial intelligence.
A user signed up for a waitlist for a humanoid robot cleaning service in San Francisco. These robots are currently piloted remotely by humans, using collected data to train their future autonomous capabilities. The cleaning process often involves the robots struggling with navigation and tasks, such as climbing stairs, due to remote operation delays. Despite imperfections, the robots successfully cleaned tasks like toilets and mopping, and the company plans to eventually deploy them for other jobs. The technology aims to eventually automate cleaning, potentially addressing labor shortages in the sector.
Aaron Nathan, CEO of Point One Navigation, has transformed GPS accuracy into a robotics platform. Nathan possesses extensive experience in sensor fusion, computer vision, navigation, and embedded systems for robotics applications. Point One Navigation is involved in this field, and they will host the FUSE 2026 physical AI event in San Francisco. Nathan has led multidisciplinary teams in finance, defense, and enterprise sectors, and founded two venture-backed startups. The discussion highlights the integration of GPS data into sophisticated robotic systems.
Boston Dynamics unveiled Stretch Gen 2, an updated autonomous mobile trailer unloading robot designed for enhanced reliability and performance. The new model features a modular design for easy part swaps, improved durability through automotive-grade connectors, and continuous rotation joints. Upgrades include a gripper that allows for multi-box picking in a single motion, reducing service time to just hours. The robot also incorporates onboard diagnostics and a continuously improving software platform that allows for remote performance enhancements. Customers have responded positively, finding the robot's user experience simple and effective for heavy lifting tasks.
Atlas uses 70 Samsung SDI 2170 cells to form a battery pack with approximately 1.3 kWh of capacity. Samsung SDI is reportedly supplying most of the initial cells, with LG Energy Solution also involved in the supply chain. The reported cells have a capacity of 5,000 mAh each. Runtime depends on various factors including task demands, payload, and computing needs, not just pack capacity. Future solutions focus on battery swapping and new chemistries to improve humanoid robot operational duration.
Most of Unitree's robot actuator production in the Hangzhou factory remains largely manual, with only a few new automated lines present. The company's 2025 filings indicate that direct labor accounted for only 8.1% of the cost of goods. The author argues that US trade barriers make factory automation more valuable to American makers than to Unitree. Automation is expected to increase significantly as AI reduces the time needed for programming and integration. Investment focus should shift to building standardized motor lines rather than solely focusing on humanoid assembly.
Canonical has developed an Ubuntu Robotics ecosystem using Snaps and Ubuntu Core, offering atomic updates and version linking for reliable robot deployment. This ecosystem includes ROS 2 extensions, observability hooks, and fleet management tools. Despite these technical advantages, there is limited discussion in the ROS 2 community about using Snaps for production robot software deployment. Alternatives like standard apt, Docker, or Yocto are currently more commonly discussed for shipping robots. The article questions if strict confinement hinders necessary hardware access or if maintenance overhead favors other methods.
Team Chiron uses autonomous drones and ground robots with AI to assess and triage wounded individuals during disasters. The system gathers data on casualties by detecting vital signs and locating injured people from a distance, aiming to help emergency responders prioritize care. This technology is being developed through participation in the DARPA Triage Challenge. Team Chiron's goal is to provide a comprehensive picture of a disaster scene to facilitate faster and more informed medical decisions. The overall aim is to advance autonomous systems that improve emergency response efficiency and save lives.
Physicl raised $10 million in seed funding to utilize its expertise in 3D rendering to improve robot training through realistic simulations. The company aims to help model designers and robot builders solve training problems by creating more accurate digital environments. Physicl offers tools, like Studio, that allow users to generate synthetic data for training robots on complex scenarios. The platform seeks to automate the process of transforming unstructured real-world information into high-fidelity 3D representations for AI. This approach aims to enable robots to better understand, interact with, and manipulate physical space.
Tau Robotics offers paid humanoid cleaning services in San Francisco, utilizing remote human operators for much of the work while collecting training data. The service charges around $60 weekly for two robots, and trials revealed heavy dependence on reliable communication and operator equipment. The robots are four and a half feet tall with cameras for remote operation and use kinematic guidance for tasks like making beds. Initial trials demonstrated operational difficulties, including struggles with stairs and equipment shutdowns dependent on human assistance. The goal is to use these real-world examples to train models for more autonomous housekeeping.
ByteDexter is a five-finger robotic research hand developed by ByteDance Seed for AI research. It is designed with an emphasis on dexterity, compact integration, and tactile sensing, featuring electric motors within the palm. The hand utilizes rigid mechanical linkages and passive four-bar linkages for finger movement. Tactile arrays measure normal force, and its performance is rated as 4/5 for degrees of freedom but only 1/5 for strength. As a prototype, it is part of the GR Dexter project, aiming to advance AI systems for robots with humanlike hands.
REX G1 is RoboScience's first wheeled humanoid robot, designed for environments like factories and homes, standing 170 cm tall and weighing 80 kg. It features an adjustable torso, two articulated arms capable of supporting a 10 kg load, and a working reach of 0.1 to 2 meters. The robot has a maximum speed of 5 km/h, operates for four hours on a charge, and includes safety brakes on all joints. Its demonstrated capabilities focus on obstacle avoidance and sorting goods, reflecting an early stage of development.
Embodied AI has introduced the AIR humanoid robot for shared factory work, with manufacturer testing currently in progress. AIR is designed to handle repetitive manual tasks that change frequently, unlike conventional automation. The robot utilizes a tendon-driven architecture with motors in the torso and wheels for stability on factory floors, prioritizing manipulation over bipedal mobility. The AI learns tasks by observing runs, allowing for human oversight during the learning phase. Physical reliability is a central design requirement, and CE marking is planned for 2027.
GoMate Mini is a humanoid robot from GAC Group, developed by Huilun Technology, featuring wheeled mobility and manipulation. Standing 170 cm tall and weighing 60 kg, it has 27 degrees of freedom and can switch between wheel configurations for operation and storage. The robot can perform tasks like obstacle avoidance, standing up after a fall, sorting goods, and assembling products. It is a fourth-generation humanoid platform with applications in automotive production and public venues. The robot is available as a prototype with a listed price of USD 40,000.
ARMAR 7 is a humanoid assistance robot developed in 2024 by the Karlsruhe Institute of Technology. It features an omnidirectional mobile base, a bowing torso, and two torque-controlled arms with integrated RGB cameras. The robot stands 175 cm tall, weighs 120 kg, and has a listed strength of 12 kg. Its demonstrated tasks include obstacle avoidance, sorting goods, and operating a dishwasher. The robot is a research platform focused on mobile manipulation in domestic assistance settings.
Honda Hand is a robotic research hand developed by Honda R&D in Japan. Its overall score is 4 out of 10, placing it in the developing category. The hand scores 3 out of 5 for Degrees of Freedom but only 1 out of 5 for Strength. This assessment highlights the need to evaluate movement and strength separately in robotic hand design. The product is a prototype with a listed price of USD 24,000.
Tianzhen Z1 is a full-size humanoid robot developed by Tlibot in Mianyang, China, measuring 167 cm tall and weighing 45 kg. It features a customizable, bionic appearance designed for public interaction, supporting up to 40 active degrees of freedom excluding hands. The robot demonstrates capabilities in obstacle avoidance for navigation and sorting goods for manipulation tasks. It is listed as a prototype with a skill score of 2 out of 10, indicating early-stage performance. The robot is available for a prototype price of USD 35,000 for exhibition, commercial, or research applications.
Gento Luna is a wheeled humanoid robot featuring a folding structure that reduces its height from 170 cm to 79 cm. It weighs 90 kg and possesses 26 degrees of freedom, with each arm having seven degrees of freedom and a 6 kg lifting capacity. The robot is equipped with 3D LiDAR for obstacle avoidance and articulated arms for sorting goods and folding clothes. It utilizes force-controlled electric joints with compliant impedance control for physical interaction. Luna is currently listed as a prototype with a skill score of 3 out of 10, targeting applications in commercial services and education.
Delta-0 is a humanoid foundation model capable of long-horizon, whole-body locomotion and manipulation using generalist policies trained on whole-body human data. It achieves this by jointly designing the policy and controller, utilizing a delta action space for seamless human intervention during execution. The model demonstrates impressive abilities in everyday tasks like operating a turntable, making the bed, and interacting with objects by coordinating movement and force. The research emphasizes that long-horizon loco-manipulation is challenging due to the coupled control, perception, and dynamic nature of the environment. Key design principles include whole-body control, execution fidelity, and compliance to enable reliable real-world deployment.