Lightberry's Lumi received a conditional FCC Covered List exemption based on a plan to move manufacturing to the United States. This exemption allows Lightberry to seek independent FCC certification before offering the robots to customers. The approval is contingent upon compliance with the onshoring plan and ongoing product vetting, with the possibility of withdrawal. Lightberry plans to begin domestic production with the robot's body, head, and hands. Achieving independent certification is the next step before off-the-shelf sales can occur.
Physical AI, combining robotics and advanced AI, is emerging as the next lucrative frontier. This sector is growing rapidly, evidenced by record funding in robotics and autonomous systems. Physical AI systems operate by perceiving environments, making real-time decisions, and adapting through real-world interaction. This advancement is fueled by maturing technologies like multimodal sensing, edge computing, and robotics foundation models. The sector is poised for significant disruption across manufacturing, logistics, and other industries.
Robotics interconnects face challenges related to signal integrity, mechanical durability, and board space due to increasing system complexity. 3M addresses these issues with a portfolio of solutions focusing on motion, density, and uptime. Their offerings include high-flex cable assemblies with validated cycle life, fast IDC termination methods, and reduced-pitch connectors to save board space. Furthermore, 3M provides solutions that enhance reliability against vibration and manage EMI/crosstalk in dense wiring. These interconnect strategies offer engineers validated options to improve the reliability and performance of next-generation robotic hardware.
DEEP Robotics humanoids, named Hangxing No.2, assisted with traffic management at the Hangzhou Marathon on November 1. These robots utilized the DR02 platform for duties including patrols, safety warnings, and public inquiries around the Huanglong Sports Center. They demonstrated the ability to perform standardized traffic gestures and broadcast safety advisories to aid roadside work. The deployment provided an opportunity to test a repeatable model for humanoid support amidst changing crowd and vehicle patterns. The actual effectiveness of this system will depend on consistent performance and clearly defined responsibilities between the robots and human officers.
1X teammates are using factory-defect NEO heads to test home integrations before complete NEO robots are available. In a demonstration, a NEO head responded to a spoken request by selecting and starting a film on a television. This informal setup showcased the potential for a home robot to understand requests and act across household devices. The demonstration only shows an interaction and does not explain the underlying technology or an imminent customer release. The company is still awaiting customer deliveries for the NEO robots, which are targeted for 2026.
Six AGIBOT G2 robots participate in a 24-hour livestream challenge set in a convenience store environment. The challenge, running from November 2nd to 3rd, involves tasks like stocking shelves, making coffee, and greeting customers during rush periods. The broadcast aims to provide an extended view of the robots operating amidst human activity. The livestream offers insight into how the robots manage routine variations and interruptions during continuous work. Complete details on task success and autonomy will be assessed based on intervention and recovery during the event.
North American companies ordered 10,989 robots valued at $723.5 million in the third quarter of 2026, marking a 25 percent increase in both units and value compared to the previous year. This growth was driven by strong demand across various sectors, including Automotive OEMs, Semiconductors/Electronics, and Life Sciences. Collaborative robot orders also saw a significant 56 percent increase during the quarter. The overall trend suggests a diversified and strengthening robotics market across different industries. Despite some economic uncertainties, the market enters the final quarter of 2026 with positive momentum.
BMW Group and Embotech have successfully moved one million vehicles using automated driving technology across four production sites in the UK and Germany. This achievement demonstrates the scalability of Automated Driving In-Plant within the BMW Group's production network. The solution supports approximately 3,500 vehicle movements daily with over 99.5 percent availability. Embotech provides the AI-based software and safety architecture necessary for this reliable operation. This deployment showcases how the technology enhances production efficiency, flexibility, and safety.
Agility Robotics is expanding the use of its Digit humanoid robot through a commercial agreement with GXO and paid programs with manufacturing customers. The company has over 30 prospective customers across various industries, including automotive and logistics. Operations at new GXO facilities are planned for early 2027 using Digit 4, transitioning to Digit 5 later. Agility is also preparing to launch Digit 5, which is designed for broader industrial workflows and is expected to be generally available in late 2027. The company is working to commercialize its robots through customer demonstrations and a cloud-based fleet management platform.
Current humanoid robots primarily use NMC-based lithium-ion batteries, balancing energy density with necessary performance criteria. Future developments are exploring solid-state batteries, which offer higher energy density and improved safety by replacing liquid electrolytes. Lithium-sulfur cells present an alternative with high theoretical energy density but face challenges with cycle life and fast charging limitations. Improvements to NMC cathodes, by increasing nickel content, can also boost performance while reducing cobalt use. Ultimately, the transition to advanced battery technologies will be driven by the need for higher energy density, safety, and longevity in humanoid applications.
TorqueAGI develops a World-Reasoning Model for industrial robots and heavy machinery by incorporating physics and geometry into a spatio-temporal architecture. This approach allows machines to adapt to changes and learn new tasks from minimal demonstrations using existing customer operational data. The model differs from vision-language-action models by building reasoning directly on spatial-temporal graphs rather than solely relying on large demonstration sets. Integration focuses on the model's intelligence, leaving hardware and sensor integration as the responsibility of the machine manufacturers. The Sovereign Physical AI architecture emphasizes running models locally to ensure operational autonomy, control, and accountability for industrial equipment.
RoboParty launched the RP1 humanoid, an open hardware and software research platform with 24 or 30 actuated degrees of freedom. Researchers have full access to mechanical designs, actuators, firmware, control algorithms, and learning tools to modify the robot. The platform supports research in gait, balance control, and embodied AI through simulation and physical deployment. It includes built-in capabilities for walking, fall recovery, and motion tracking, along with tools like VR teleoperation. This open approach allows teams to inspect and adjust all components for practical debugging advantages.
CR Hand is a compact robotic hand from Curious Robotics, designed for education and research. It costs USD 1,500 and is constructed from FDM printed plastic with integrated servo motors and a controller. The hand has a Humanoid.Guide Hand Score of 3 out of 10, reflecting developing dexterity and lower strength. It serves as a platform to examine the balance between dexterity, strength, weight, and cost in robotic hand design. The product is available for purchase, emphasizing an affordable, practical design for educational use.
Elf Xuan 2.0 is an upper body humanoid developed by AheadForm for expressive interaction and stationary public presentation. It features an elven design with lifelike silicone skin, articulated arms, and a movable waist with sophisticated facial expressions. The robot is 165 cm tall and weighs 50 kg, utilizing brushless micro motors for facial movements. Its physical capabilities are rated very low, placing it as an expressive humanoid rather than a broadly capable mobile worker. As a prototype, it is priced at USD 55,000, emphasizing its visual presentation over advanced physical autonomy.
The YMBot D4 is an industrial mobile humanoid robot featuring a humanoid upper body and a folding wheeled chassis, designed for factory and logistics environments. It possesses two articulated arms with five-fingered hands for sorting goods and assembling products, supported by obstacle avoidance for navigation. Standing 165 cm tall and weighing 100 kg, it has a maximum speed of 5.4 km/h and a 10 kg strength capacity. Despite its combined functions, its skill score is low, indicating its development level is in the developing category. The robot supports ROS 2 communication and is offered as a prototype with a listed price of USD 55,000.
Meta has reportedly built a humanoid prototype named Max in collaboration with the Chinese robotics company Fourier. This prototype is intended as a physical embodiment of Meta’s Muse agent, focusing on household tasks. The robot was developed in about three months and weighs approximately 106 pounds (48 kg). Meta plans to build a second version but has no plans to sell these units, positioning it as an internal development platform. The partnership leverages Fourier's expertise in robotics hardware to support Meta's AI and software development.
Billi is a wheeled, hybrid humanoid robot developed by NEUBILITY for industrial logistics and material handling. It features a mobile base, a humanlike upper body, and two arms with five-fingered hands, offering a rotating and adjustable upper body. The robot has a maximum speed of 6 km/h, a weight of 110 kg, and capabilities in camera-based perception for navigation and obstacle avoidance. Its performance ratings place it in an early stage for general-purpose systems, with navigation and manipulation rated 1 out of 5. Billi functions as an early prototype, focusing on a mobile platform for sorting and navigation in industrial settings.
YMBot C is a full-size bipedal humanoid, 174 cm tall and weighing 75 kg, developed by Jiangsu Yunmu Intelligent Manufacturing Technology Co., Ltd. It is designed for cultural tourism and exhibitions, featuring a humanlike exterior with two arms and hands. Its confirmed capability is obstacle avoidance for navigation, with low scores for overall skill, navigation, and manipulation. The robot is currently a prototype with a listed price of USD 60,000.
RealBot 01 is a wheeled humanoid robot from RealMan Robotics, measuring 168 cm tall and weighing 125 kg, capable of traveling up to 5.4 km/h. It features two articulated arms with a combined payload capacity of 10 kg, and a four-hour runtime per charge. The robot is modular, allowing for integration with other platforms, and runs on Ubuntu 22.04 LTS with ROS 2 Humble. Documented capabilities include obstacle avoidance, sorting goods, and product assembly, focusing on combined navigation and manipulation. The platform targets embodied AI and industrial automation markets and is listed for a price of USD 30,000.
Imitation learning policies often fail in real-world robotic manipulation because conventional optimization signals do not align with actual success. Evaluation of 281 policies using real-robot trials on grasping and insertion revealed that observation and action representations were dominant factors. Specific representation choices, such as region-of-interest cropping for point-cloud policies, significantly improved performance. Furthermore, absolute action prediction rather than relative prediction boosted insertion success for vision-language-action policies. Representation choices are critical determinants of imitation learning performance, though offline metrics alone are insufficient.
FlapSim is a modular MuJoCo simulation platform designed for flapping-wing robotics by combining robot descriptions, wing geometry, and composable aerodynamic models. It evaluates five reduced-order aerodynamic configurations against a high-fidelity $50\,\mu\text{s}$ blade-element reference across various flight conditions. Incorporating multirate synchronization, inflow/gust coupling, and harmonic synthesis significantly improves simulation fidelity. The results indicate that timing, inflow, geometry, and harmonic representation are critical factors in the accuracy-computation tradeoff of the reduced-order model. Ultimately, the work characterizes the necessary trade-offs for simulating flapping-wing dynamics effectively.
End-to-end policies require safety filters, which typically maintain a fixed level of conservativeness regardless of the policy's actual competence. The authors propose a method to modulate this filtering based on the policy's confidence level at runtime. Inside the operational design domain (ODD), high confidence allows for performance gains, while low confidence introduces a margin to flag where the filter may be failing. This approach combines the strengths of both permissive and conservative filters. Preliminary simulations suggest this confidence-modulated filtering effectively combines these filter types.
Vision-language-action models show strong capabilities in robotic manipulation but action representations near the output do not fully preserve task-relevant visual grounding. STAR introduces a training-free inference method for pre-trained VLAs by applying an exponential moving average over late-layer residual updates. This method allows later action representations to inherit object-aware update directions from earlier layers without model updates. STAR also uses proprioceptive states to adjust mixing strength, employing conservative mixing when execution dynamics change rapidly. Experiments demonstrate that STAR improves late-layer object awareness and increases average task success rates by 4.0 percentage points.
A framework called Contact Imitation Interface (CIMI) is presented to transfer dexterous skills from human demonstrations to robots. CIMI utilizes contact-grounded demonstration and learning interfaces to address challenges in producing robot-transferable actions. The demonstration interface combines shared contact geometry and fingertip RGB views with contact-alignment-first retargeting. The learning interface employs a hierarchical policy that predicts contact-surface trajectories and coordinates finger/wrist commands. This method improves the mean final-stage success by at least 50 percentage points across six hand-task pairs.
Veloxity is an open-source, memory-safe embedded UAV autopilot developed using the Rust programming language. It aims to address the lack of memory-safe practices in the aerospace domain for UAV firmware. The project provides a minimal, modular foundation to lower the barrier for using memory-safe languages in modern ecosystems. Performance comparisons with C++ autopilots in simulation and flight tests demonstrate functional parity. These results suggest that memory-safe languages are viable for reliable and performant autonomous aerospace systems.
End-to-end self-driving models lack transparency despite strong planning performance. Mechanistic interpretability is applied using Sparse Autoencoders (SAEs) on a single-camera path planning model's internal activations. Unsupervised clustering of SAE latents reveals interpretable structures related to route intent, speed, and braking, accurately recovering route intent. These SAE features are found to be visually grounded and systematically respond to scene changes. Finally, using these representations in an adapter improved trajectory selection error, confirming their utility for interpretability.
Safe navigation for heterogeneous multi-robot teams is complicated by differing braking and turning capabilities and the need for coordinated control actions. The Capability-Aware Backup-Recoverability Filter (CA-BRF) is proposed to address this by ensuring navigation actions preserve access to certified recovery maneuvers. It constructs backup-recoverable sets based on pairwise safety and robot constraints to define capability-dependent recovery certificates. CA-BRF filters candidate controls by selecting admissible interpolations, preserving recoverability while satisfying all pairwise recovery conditions. This method establishes recursive feasibility and safety for the system during complex interactions involving up to 30 heterogeneous robots.