Trillions of dollars have been wiped from Chinese stock markets as the race for AI supremacy with the US causes concern. This decline is linked to a lack of profits and the threat posed by Western countries imposing restrictions on Chinese technology. Despite heavy government investment in AI, weak domestic consumer demand has hindered profitability for many tech companies. China is focusing on embodied AI and robotics to advance technology, but consumer demand for these products is currently low. Concerns persist over the sustainability of the AI boom amid these economic and geopolitical pressures.
The Royal Air Force chief advocates for an "autonomy, not flesh" future in warfare, suggesting AI-enabled drones will reduce the need for direct human involvement. Typhoon fighters will be accompanied by four AI drones capable of long-range patrolling. This technological shift aims to move military roles "towards the outside" and ease recruitment pressures. The drones will provide a force multiplier, placing sensors and weapons within contested airspace. This initiative aligns with the goal of the RAF to be ready to fight against Russia by 2030.
Xpanner introduced the X1 Piling R6, an automated robot designed to autonomously install posts for utility-scale solar and battery energy storage sites. The system uses high-precision GNSS and onboard sensors to ensure accurate pile placement, supervised by a single operator. The robot is powered by a 97 kW engine and hydraulic hammer capable of driving piles up to 23 feet long. Project teams can use Xpanner Connect to manage schedules and monitor installation data remotely. The business model is a mixed Capex/SaaS subscription, with Xpanner handling initial setup and support.
WorldOracle is a benchmark evaluating action-conditioned world models by comparing their predictions against real-world robot executions. It assesses how accurately models predict object positions and movements during manipulation tasks, noting that arm motion is often predicted well but object states frequently deviate from reality. The evaluation also tests model performance over different time horizons, showing that accuracy often degrades over complete operation sequences. The platform measures prediction errors using metrics like Physical State Correctness and Interaction Spatial Alignment, which score alignment with real outcomes. Developers can participate by submitting predictions from open-weight or closed-source models for iterative improvement.
Lars Moravy suggested that Tesla's Optimus could assist with charging Cybercab vehicles by interacting with existing charging infrastructure. This proposal focuses on Optimus bridging the gap in physical interaction required for connecting to established charging systems. Moravy expressed hope for this capability but did not provide a firm deployment date or operational commitment. The discussion highlighted that this role involves physical interaction with built infrastructure rather than fully automating the entire charging process. The article also briefly mentioned Optimus's battery size and production status.
Boston Dynamics has solicited Korean suppliers to prepare for purchase orders for 10,000 humanoid robots, according to a KED Global report. Hyundai Mobis is positioned as a central supplier, providing most actuators and collaborating with Korean battery and camera manufacturers. The reported supplier selection prioritizes cost control and reducing dependence on Chinese components. Final quantities and pricing are undecided, with initial orders potentially appearing in late 2027 or early 2028. This supply chain effort leverages Hyundai's automotive manufacturing experience for Atlas production.
Boston Dynamics is proceeding with a deal to supply modified Spot robots to U.S. Immigration and Customs Enforcement despite an employee petition urging them to stop. The dispute centers on the extent of responsibility robot makers hold over customer applications and acceptable uses. Employees argued the association with ICE could damage the company's reputation and impact recruitment. Boston Dynamics management confirmed the sale, emphasizing restrictions against weaponization and harm. The situation highlights the challenge for physical AI companies in establishing guidelines for customer deployment beyond safety concerns.
Figure is developing Optimus Gen 3, with a likely reveal in late 2026 or early 2027, focusing on human-like hands and simplified manufacturing. Tesla is building Optimus Gen 3 robots internally, with customer deliveries potentially toward the end of 2027. Apptronik is working on its next Apollo successor, with an estimated reveal in 2027. Other leading humanoid makers include Boston Dynamics (Atlas), Agility Robotics (Digit 5), and 1X (NEO). None of these groups have publicly announced firm reveal dates for their next-generation robots.
A field test investigated the RTK fix quality for outdoor robots operating under tree canopies. The test compared performance under open sky, under trees with rain, and under trees with network RTK corrections. In long-baseline sessions, most receivers achieved 100% fix status, though some dropped to 83.3% or 88.5% fixed under tree conditions. The study cautions that fixed flags do not prove accuracy due to varying conditions and lack of ground truth. The author recommends a specific test design for evaluating receivers in tree environments.
Comau is integrating Nvidia technologies to advance the development of AI-powered robotics and automation for industrial use. This collaboration involves powering industrial robot controllers with Nvidia Jetson edge compute and integrating the robot portfolio into Nvidia Isaac Sim for simulation and testing. Comau is also using the Nvidia software stack to optimize vision features and control strategies for intralogistics products. This work includes training AI models using reinforcement learning for tasks like palletizing and depalletizing within these simulated environments. Ultimately, this approach aims to accelerate the deployment of smarter, more autonomous robotic solutions in real-world industrial settings.
Edeka expanded its automated logistics at its Nordfrische distribution center by integrating two Knapp Pick-it-Easy robots for meat picking. These AI-powered robots operate 24/7, autonomously recognizing and handling various fresh meat products with high efficiency. The system supports around 450 picks per robot per hour while also scanning barcodes and recording weights to improve process quality. This expansion builds upon a long-standing partnership and utilizes Knapp's KiSoft software for intelligent control across all logistics processes. The automation enhances supply chain resilience and allows employees to be relieved from strenuous tasks.
Omron will acquire US industrial computer maker OnLogic and its parent company Interlogic for $675 million. This acquisition supports Omron's strategy to advance its industrial automation business into manufacturing digitalization and physical AI. OnLogic designs configurable industrial computers for automation, edge AI, and IoT applications. The transaction is expected to close around January 2027. Omron plans to integrate OnLogic's hardware with its own technologies to create new data- and AI-intensive industrial solutions.
Qualcomm agreed to acquire robotics software company PickNik to enhance its open-source robotics and physical AI capabilities. This acquisition aims to simplify the transition for developers from AI models to production robotics systems. MoveIt, an open-source framework providing motion planning and manipulation capabilities, will remain open source with community support. Qualcomm plans to integrate MoveIt and PickNik's commercial platform with its Dragonwing robotics platforms and Arduino boards. The combined effort seeks to connect AI models with motion planning for advanced robotic applications.
Charl-Pol and Norway's Storvik have formed Alumotion, a joint venture focused on automation and robotics for Canadian aluminum smelters. This partnership aims to boost operational efficiency and address workforce challenges by reducing exposure to high-risk tasks. Alumotion will specifically focus on using robotics for anode assembly repair, a task currently done manually. The collaboration leverages Charl-Pol's local smelter knowledge and Storvik's European automation expertise. This alliance expands Charl-Pol's international collaborations in the aluminum sector.
Orbbec Gemini 330 series active stereo cameras use the proprietary MX6800 ASIC to compute depth and align it to color internally. These cameras operate active and passive stereo simultaneously, allowing for performance across various lighting conditions. The in-camera hardware handles depth generation and depth-to-color alignment, reducing the computational load on the host. The family includes models like the Gemini 335 for compact setups and the Gemini 335L for long-baseline applications requiring greater range and accuracy. This architecture targets robotics and automation by offering integrated, robust, and low-latency depth sensing.
Bedrock Robotics secured $600 million in a Series C financing to scale autonomous excavators across the US. This funding, co-led by CapitalG and Emergence Capital, supports expanding the deployment of autonomous machines on commercial construction sites. The company has already demonstrated fully autonomous excavators performing paid work on live job sites. Bedrock is focused on building the necessary field, commercial, and operational infrastructure to make autonomy a dependable part of construction. New funding will allow Bedrock to expand deployments, strengthen field operations, and work with more partners to integrate intelligent systems into construction.
4NE1 Gen 4 is a 180 cm tall, 65 kg humanoid robot developed by NEURA Robotics in Germany for physical work alongside people. It features 55 overall degrees of freedom, with enhanced capabilities in the fourth generation, including lighter textile skin and more precise hands. The robot can perform tasks such as sorting goods, assembling products, and managing household chores like folding clothes and emptying a dishwasher. It utilizes onboard NEURON AI processing with multi-sensory feedback for interaction and navigation, supporting obstacle avoidance. The robot is currently in production with a price of USD 70,000, emphasizing integrated industrial and household manipulation.
The ZWHAND B20 is a robotic hand from Shenzhen ZWHAND Technology Co., Ltd., emphasizing dexterity and controlled manipulation. It received a Humanoid.Guide Hand Score of 6 out of 10, reflecting a capable or mature performance level. The rating breakdown shows a Degrees of Freedom rating of 4 out of 5, which is stronger than its Strength rating of 2 out of 5. This separation allows users to prioritize dexterity over gripping force when evaluating the hand. The B20 is currently in production, costs USD 6,500, weighs 0.6 kilograms, and is manufactured in China.
GXO Logistics will expand its Agility Digit deployment to two new warehouses in North Carolina and Tennessee, starting installations in early 2027. The initial deployment will use Digit 4 for e-commerce fulfillment, with plans to adopt Digit 5 later. This expansion builds upon a pilot program that began at the Flowery Branch, Georgia facility. The expansion demonstrates customer value, even as GXO noted a lack of return on investment for humanoid robots. GXO will begin with Digit 4 and migrate to the newer Digit 5 once it is commercially available.
StarBot Nova is a wheeled service humanoid developed by StarBot USA for restaurant, hospitality, and retail environments. It measures 152 cm tall and weighs 32 kilograms, featuring a differential drive base and two articulated arms. The robot can move at 4.3 km/h, has a strength rating of 5 kilograms, and a 10-hour runtime per charge. Its confirmed capabilities include obstacle avoidance and sorting goods using Intel RealSense and RGB cameras. Nova is a prototype with a published price of USD 20,000, positioning it as an early platform for commercial mobile sorting tasks.
MagicBot Gen1 is a 174 cm tall, 70 kg humanoid robot developed by MagicLab in Wuxi, China, designed for research, manufacturing, and service. It features bipedal locomotion, articulated arms, and optional tactile hands, with an 8-core CPU and a 100 TOPS AI processor for onboard computing. The robot's supported capabilities focus on sorting goods and assembling products, using LiDAR and depth vision for obstacle avoidance. It is listed as in production with a price of USD 50,000 and supports connectivity standards like 5G and WiFi 6. Currently, its skill score places it in the developing category, suggesting a focus on structured trials rather than advanced autonomy.
The ZWHAND A17 is a compact robotic hand manufactured by Shenzhen ZWHAND Technology Co., Ltd. It has a Humanoid.Guide Hand Score of 6 out of 10, indicating a capable or mature category. Its rating shows a higher dexterity score (4 out of 5) than its strength score (2 out of 5), suggesting a focus on dexterity over raw strength. The hand measures 244mm long and weighs 0.85 kilograms, closely resembling a human hand. Listed at USD 5,000, it fits between the B06 and B20 models in the manufacturer's product family.
The Linker Hand L30 Pro is a robotic hand from LINKERBOT with a Humanoid.Guide Hand Score of 7 out of 10, indicating a capable or mature level. Its rating profile emphasizes dexterity, achieving a perfect 5 out of 5 for Degrees of Freedom, while its Strength rating is 2 out of 5. This highlights a balance prioritizing movement versatility over raw gripping force. The evaluation suggests that buyers should consider the importance of dexterity versus force for their specific applications. LINKERBOT lists the L30 Pro as in production, with pricing available upon inquiry.
RAIWAY is a large-wheeled bipedal robot featuring 20-inch wheels and a contact-informed control approach for locomotion. The system manages contact-dominated rolling locomotion by distilling multi-contact events into a latent representation from simulation data. The controller infers this representation online using proprioception, eliminating the need for direct contact sensing during operation. Experiments showed RAIWAY successfully traversed diverse challenging terrains, including steep slopes and rocky surfaces. The proposed controller outperformed baselines in both indoor and outdoor testing environments.
Rapid illumination changes can disrupt Visual SLAM by saturating images before exposure adaptation occurs, removing necessary visual correspondences. The proposed method, B-Adapt, acquires adaptive exposure brackets to cover the scene dynamic range while maintaining photometric overlap. BVSLAM then directly uses this multi-exposure sequence as constraints in a factor graph state estimator, bypassing HDR merging. Evaluation showed that combining these methods reduced state estimation error by 78% compared to using standard Automatic Exposure images during tunnel-exit transitions. This demonstrates that bracketing improves visual pose estimation robustness to illumination disturbances.
The research proposes a modular perception-prediction-planning pipeline to ensure interpretable collision avoidance and feasibility guarantees for autonomous vehicles. This pipeline propagates uncertainty as structured quantities across modules, decoupling safety guarantees from solver feasibility. It resolves collision probabilities upstream to construct convex safe sets for the planner. Experimental results show the pipeline achieves zero collisions and infeasibility at real-time speeds, outperforming baseline methods. The approach is validated on physical platforms and demonstrates real-time viability.
Deep learning in robotics often results in actions without a true understanding of the underlying reasons, leading to performance and trustworthiness limitations. Causal-based approaches are proposed to provide the necessary reasoning architecture for robots to understand and explain their behavior. These explanations, when validated, allow robots to incorporate the understanding into their knowledge base for future improvements. The review examines recent advances in creating self-aware and self-evolving robots that offer these explanations to humans. Ultimately, the goal is to enhance robot capabilities and build human trust through causal reasoning.
ORBIT is a modular, multi-finger robotic hand platform designed to allow for the systematic comparison of different hand morphologies. Its core feature is a rotating knuckle mechanism enabling quick reconfiguration of finger layouts without disassembly. The platform demonstrates the ability to perform tasks difficult for fixed-knuckle designs, such as reorienting cube faces and precision pinching. Pilot grasping studies compared anthropomorphic, isometric, and Barrett-style finger layouts across various objects. These results establish ORBIT as a viable platform for directly comparing multi-fingered hand morphologies.
Neuromorphic decision dynamics (NDD) is introduced as a decentralized, event-based controller for safe navigation in congested environments. NDD ensures safety, liveness, robustness, and scalability by making safe-and-live deconfliction an emergent property of local agent interactions. Agents modulate speed based on local motion, allowing them to self-organize into conflict-resolving groups. These groups emerge and stabilize when safety-critical conflicts trigger coordinated resolution, guaranteeing collision avoidance without deadlock or livelock. Explicit conditions ensure collision avoidance and traversal-order formation for finite, heterogeneous agent populations.
Tiny Recursive Policy (TRP) is a parameter-efficient behavior-cloning policy that incorporates persistent temporal state with recursive latent refinement for visuomotor control. TRP updates high- and low-level latent states repeatedly at each control step using a shared module before generating actions. The method evaluates well on benchmarks like LIBERO-100 and MetaWorld ML45, showing competitive performance in multi-task and few-shot settings. The study indicates that reliance on persistent carry increases with temporal-context demand, and training with carry perturbations enhances robustness. Ablations show that the benefit of the persistent state interacts strongly with the amount of within-step computation.
The research develops a framework to monitor the recoverability of humanoid policies during deployment to anticipate unrecoverable falls. It addresses the challenge that existing methods lack sufficient training data for the full recoverability envelope, leading to optimistic hallucinations for out-of-distribution inputs. The proposed solution uses a unified framework that alternates between value learning and active exploration, reframing data exploration as a reach-avoid problem. This method systematically discovers unseen states while avoiding confidently unrecoverable regions, supported by combined TD and Monte Carlo objectives. The approach successfully improves monitor coverage and accuracy on benchmarks and transfers to real-world humanoid hardware.
MOTIO is an agentic skill-engineering system designed for reality-grounded humanoid loco-manipulation. It synthesizes human action videos, 3D assets, motion references, and robot skills into executable programs. The system uses in-context reasoning to construct staged interactions and refines them via physics execution and human validation. Agent-generated skills supervise distillation into a shared reference policy, enabling task completion without constant expert intervention. This approach successfully prepares long-horizon skills, such as bottle disposal, and reduces physical residual errors.