Industrial robot installations grew to over five million globally in 2026, with China leading in installations. Research is advancing anthropomorphic robotic hands capable of self-supported locomotion and manipulation through reinforcement learning. Investigations in India reveal women are paid to record household tasks to train robots on real-world manipulation. Discussions are ongoing regarding the regulation and potential prohibition of autonomous weapons systems. Studies suggest that perceived fairness, both distributive and informational, can improve acceptance of security robots.
advasolutions is implementing a tote-based storage system for the Polish drugstore chain Hebe. This project involves retrofitting an existing building with 66 autonomous advaShuttles to store approximately 20,000 totes at ten meters high. The implementation will occur in phases during ongoing operations, with a go-live scheduled for early 2028, integrated with SAP EWM. The system features automated replenishment and incorporates conveyor technology for discrepancy handling. This solution demonstrates how modular automation can be integrated into live operational environments without major structural changes.
Europe is emerging as a crucial battleground for the global robotaxi race due to its relative openness compared to the US and China. This fragmentation creates opportunities for robotaxi companies seeking an early market foothold by allowing direct competition between Chinese and Western firms. Success in Europe will depend on vehicle economics, fleet utilization, regulation, consumer trust, and the ability to operate profitably at scale. Markets like Spain and the UK offer distinct opportunities based on regulatory timing and market size. Ultimately, fleet operators who can optimize vehicle utilization and costs are expected to be the long-term winners.
Compressed air dictates the performance of automated lines because the quality and pressure of the supply directly affect the operation of pneumatic components. Inadequate air delivery causes cycle times to stretch and parts to slip, often leading to maintenance issues that are hard to trace. Leaks, corrosion from moisture, and material choices in the piping network significantly impact long-term operational reliability. Effective distribution requires careful leak detection and material selection to minimize production losses. Modern designs favor flexible loop systems and quick-connect fittings to accommodate plant changes while maintaining system integrity.
Teradyne Robotics partnered with Sereact to advance physical AI solutions for retail and logistics. Sereact's Cortex platform runs Universal Robots robot arms to automate unstructured tasks like handling returns and picking. This combined technology handles mixed goods and varied packaging, allowing a single "brain" to manage various tasks across robots. The joint solution is already deployed in Zalando's fulfillment centers in Europe for processing returned fashion items. The collaboration aims to scale these AI-powered automation solutions across various industries globally.
Skyfly Technologies secured a $10 million seed round led by HICO Investment Group and Archangel to advance its electric aircraft. The funding will be used to complete the certification of the Axe, a two-seat eVTOL aircraft, and expand production. The Axe is designed to meet updated FAA Light Sport Aircraft rules, offering a cost-effective and aerodynamically efficient design. Skyfly is also developing the Longbow, a hybrid-electric unmanned aircraft for defense and logistics applications. The company's prototype has flown extensively and has a strong pipeline of potential customers.
Schaeffler opened a state-of-the-art lighthouse factory in Taicang, China, dedicated to humanoid robotics. This facility features an intelligent assembly line and a training center to facilitate the mass production of humanoid robot components. The factory demonstrates a closed manufacturing loop where humanoids assemble key joint modules for other humanoids. It combines manufacturing expertise with physical AI and digital twins for testing and optimizing production. This initiative aims to drive the adoption of humanoids in industry by accelerating the development of the humanoid robotics ecosystem.
beRobox introduced the "3D Dynamic Safety" module for PALTZ robotic palletizers, creating a dynamic safety zone that moves with the robot. This system uses Sonair's safety-certified ADAR One 3D ultrasonic sensor to provide real-time, three-dimensional detection of people and objects. This allows the safety system to adapt to the dynamic palletizing workspace, enabling safe operator access during operations. The technology allows for more selective reactions to operator interactions without stopping the entire cell. Furthermore, the system maintains presence detection around the robot, preventing unintended restarts when an area is occupied.
An individual wondered if a mane of bristles could serve as lightweight, impact-resistant armor for a combat robot. This experiment was conducted on a robot competing in a plastic antweight division, where the concept seemed promising. The robot's hairy design withstood impacts that broke the outer shells of previous robots. This suggests the idea may have merit for combat robotics designs. Furthermore, 3D printing facilitates the experimentation with unorthodox robot designs.
TRON 2 is a modular embodied robot from LimX Dynamics designed for research and prototyping, offering configurable forms like dual arm, bipedal, or wheeled legs. It measures 132 cm, weighs 35 kg, and has a maximum speed of 18 km/h, supporting features such as obstacle avoidance, stair walking, and goods sorting. The robot's skill score is 4/10, indicating it is currently stronger in mobility than manipulation. Developed with a Linux-based system supporting ROS and features for VR teleoperation, it is available for USD 37,700.
StarBot Astra is a 152 cm tall, 34 kg humanoid research platform designed for embodied robotics investigation in supervised settings. It features 23 degrees of freedom, electric joints, and hands with five fingers, running a Linux and ROS software environment on an NVIDIA Jetson Orin system. The platform supports research in obstacle avoidance and sorting goods, which serve as focused experimental areas. Its low Humanoid Guide score and preliminary ratings indicate it is a preliminary research system, not a mature general-purpose worker. Astra is available for preorder at a starting price of $9,999 for academic and research institutions.
Toyota plans to invest $6.4 billion annually in factory automation, deploying 400,000 robots to commercialize robotics research. The humanoid strategy focuses on automating factory work and addressing labor shortages rather than pursuing an external robot business first. Toyota's internally developed Eley humanoid learns tasks by observing factory workers using an in-house foundation model. The company emphasizes integrating robots, AI, and production processes into a unified operating system as a competitive advantage. Deployment priority is set for internal factories, with projections estimating Toyota could capture a small segment of the humanoid market by 2035.
The Linker Hand L20 is a robotic hand from LinkerBot, featuring five fingers and a weight of approximately 1.2 kg. Its Humanoid.Guide Hand Score is 6 out of 10, indicating a preference for dexterity over strength in its design. The mechanical design utilizes brushless motors and a rigid linkage transmission for movement. It operates on a voltage range of 24 to 48 volts with CAN or RS485 interfaces. The hand is listed as in production with an advertised price of USD 4,200.
Tau Humanoid is a 132 cm tall, 35 kg humanoid robot developed by Tau Robotics for home and office cleaning. It is built with aluminum alloy and engineering plastics and features capabilities like obstacle avoidance, sorting goods, and folding clothes. The robot operates with AI assistance and remote human guidance rather than full autonomy. Its skill scores indicate stronger manipulation performance than navigation ability. Tau Robotics offers the robot through a cleaning service, requiring remote human supervision for tasks.
The RH5MK1 is a dexterous robotic hand manufactured by Inspire Robotics in China, earning an overall score of 9 out of 10. It rates 5 out of 5 for Degrees of Freedom and 4 out of 5 for Strength, highlighting its dexterity. This combination positions the hand as a notable option for advanced robotic hands. The hand is currently in production and is advertised at a price of USD 15,000. The dual ratings allow for useful comparison between dexterity and strength in robotic hand selection.
An ETH Zurich team developed a single policy enabling a Unitree G1 to walk, climb, and crawl without manually scripting transitions between these movements. They trained the system by using 13.2 minutes of human motion data to learn contact patterns. Reinforcement learning in simulation was then used to allow the robot to adapt to new terrains. During operation, onboard sensing and a goal guide the robot's movements without requiring a reference motion for transitions. The team demonstrated climbing, including a ladder recovery, using the learned policy rather than explicitly programmed sequences.
AgiBot Expedition A3 humanoids achieved a Guinness World Record for the most humanoid robots dancing simultaneously with 110 machines in Jinan, China, on October 24, 2026. The robots coordinated their routines using AgiBot’s LinkBots cluster control and ultra wideband positioning technology, achieving centimeter-level accuracy. The performance included complex maneuvers such as repositioning, jumps, and flips. The record demonstrates fleet coordination rather than industrial productivity, highlighting the challenge of maintaining synchronized motion among numerous mobile machines. Future applications for these robots will focus on cultural performances, public services, and education.
RepresentNav is a navigation foundation model that constructs a self-constructed representation space from vision foundation models. It uses Representation Residuals, a transformer, to aggregate features from multiple vision models to understand navigation needs. The model also generates "dreams" through a latent chain-of-thought to predict scene representations for future moments. This dream generation is trained jointly with the navigation policy via a joint-embedding predictive architecture (JEPA). RepresentNav achieves state-of-the-art results on several benchmarks and improves physical robot performance amid clutter.
KiRoSe is a framework that utilizes robot kinematic data for efficiently segmenting robot demonstrations. It proposes Sparse Motion Tokens (SMTs), a discrete vocabulary to compactly represent recurring kinematic events, inspired by entropy coding. SMTs are learned by encoding trajectories with temporal supervision and quantizing features. Evaluation shows KiRoSe significantly outperforms baseline methods in F1 quality segmentation. Furthermore, the method can track task progress and detect action reversals on unseen robots.
SEVI is a framework designed for reliability-aware safety intervention during real-world reinforcement learning to address unsafe interactions. It introduces an Evidential Safety Critic that estimates future constraint-violation risk and epistemic uncertainty in a single forward pass. During online learning, SEVI checks actions before execution and overrides them with an offline-trained recovery policy if risk is too high. The method was evaluated on simulated and real-world manipulation tasks, demonstrating superior safety prediction and reduced unsafe events compared to other methods. SEVI successfully reduces unsafe events by 13.4 to 57.1 percent while maintaining low-latency execution.
WorldMirror is an agentic framework that generates interactive digital twins from in-the-wild robot demonstrations. It treats real-to-sim construction as iteratively refining an executable world program through agent planning and tool orchestration. The system utilizes a camera-pose-anchored scene representation to organize spatial context, geometry, and interactive assets. An agent refines reconstructions by diagnosing and correcting errors to ensure physically consistent interactions and trajectory replay. Experiments show that WorldMirror can create visually faithful and geometrically consistent interactive worlds while maintaining the physical validity of recorded robot trajectories.
FORCA is a coding agent designed for contact-rich robot manipulation that translates visual observations and force/torque measurements into Python programs. It inspects force records and visual outcomes between executions to diagnose failures and revise control logic. The agent achieves this without requiring demonstrations, task-specific training, or simulation. Tested across five real-robot tasks, FORCA improves manipulation accuracy and adapts to changes in object properties. Compared to a vision-only agent, FORCA enhances grasp accuracy and reduces contact loss during tasks like wiping and peeling.
EgoForesight is a vision-language-action framework designed to scale policy learning despite noisy egocentric actions. It learns jointly from robot demonstrations and egocentric videos by mitigating the influence of unreliable action labels. The method employs a joint vision-action flow and reliability-conditioned corruption to manage noisy supervision. Furthermore, it utilizes privileged-future representation distillation from demonstrated outcomes to provide complementary supervision. This approach results in steeper scaling slopes, improving performance in real-world and simulated dexterous manipulation tasks.
The paper introduces BM2PC, a Bayesian Multimodal MPC framework for the agile manipulation of deformable linear objects (DLOs). It formulates multimodal planning as a set optimization problem to balance control quality and behavioral diversity. BM2PC efficiently updates multiple promising control modes with low computational overhead through an incremental approach. The framework incorporates an execution dynamics model to compensate for real-world execution mismatches during manipulation. Experiments show BM2PC achieves accurate long-horizon tracking and efficient obstacle avoidance in simulation and hardware.