Federal safety regulators are investigating why Waymo robotaxis continued to illegally pass school buses after a software fix. The National Highway Traffic Safety Administration (NHTSA) noted that most reported incidents occurred after the company deployed the update intended to resolve the issue. An initial investigation was triggered after footage showed a Waymo vehicle crossing in front of a stopped school bus in Atlanta. Waymo stated they have made improvements and received no citations since March 2026. NHTSA has opened a "recall query" to assess the adequacy of the implemented remedy.
China blocked SoftBank Group's $5.4 billion acquisition of ABB's robotics business due to national security concerns. This decision stemmed from a review by China's Office of the Working Mechanism for Foreign Investment Security Review. The acquisition was significant as it would have given SoftBank control of a major industrial robotics manufacturer. Concerns also focused on the potential transfer of robotics data and intellectual property related to AI training. Despite the block, ABB and SoftBank are reportedly still working on discussions regarding a potential sale.
Write performance optimizations have been implemented in the mcap libraries. rosbag2 in ROS 2 Rolling has been updated to utilize these new libraries. Consequently, writing bags should now be faster than previously. Further details on these improvements are available in the Foxglove documentation. The author invites questions and suggestions regarding these updates.
A neural network input module named PRISM improved simulated humanoid walking by enabling models to complete 92.5% of 24-second trials without falling. PRISM achieved this by learning multiplicative relationships among observations, past actions, and commands, offering a more explicit representation of physical interactions. This architectural improvement resulted in superior performance compared to conventional multilayer perceptrons, even when friction, mass, and payload were altered. The module's effectiveness was shown across various physical conditions without requiring retraining. PRISM's learned representations were found to make quantities like joint power and contact forces more predictable.
Vision AI and robotics are utilized to sort mixed recyclables. These technologies operate on conveyor belts to identify valuable materials from used clothing, bottles, and other mixed waste. This process diverts thousands of tons of materials away from landfills. The AI and robotic systems efficiently pick out recoverable items. Consequently, this method improves the recycling process.
NVIDIA Alpamayo 2 Edge is a 2B-parameter vision-language-action model designed for real-time operation on NVIDIA DRIVE AGX Thor. It generates Chain-of-Causation reasoning and driving trajectories using surround-camera observations and navigation inputs, adapting reasoning only when necessary. The model is improved through closed-loop reinforcement learning in AlpaGym and evaluated using AlpaJudge for reasoning quality. Development workflows use AlpaJudge to select training examples and generate reasoning-based rewards for post-training. This process creates a repeatable loop to bring reasoning-based driving to the edge while managing compute and latency constraints.
Dynamic Creatures partnered with Sixth Wall Studio to integrate Hasbro characters with interactive robots. The goal is to create robots that can greet guests, share adventures, and remember them in entertainment and hospitality venues. This collaboration merges Sixth Wall Studio's character authoring with Dynamic Creatures' expressive robots to give characters physical presence. The robots will translate a character's personality, voice, and story into physical movements and interactions. This technology aims to enhance guest experiences in places like theme parks, resorts, and hotels through personalized encounters.
Factories increasingly use robots, making predictive maintenance crucial for ensuring uptime and productivity. Robotic systems generate data like vibration, temperature, and current, which can signal developing faults. Predictive maintenance uses this data, along with context from the workcell, to detect changes before a failure occurs. This allows maintenance teams to schedule timely inspections, reduce unplanned downtime, and improve asset life. Ultimately, predictive maintenance enables proactive repairs, leading to greater operational stability and efficiency in robotic environments.
Intel and Altera are collaborating to develop flexible, real-time sensor-bridge reference systems for edge and physical AI applications. This collaboration combines Intel's compute and software with Altera's Field-Programmable Gate Arrays (FPGAs) to manage complex data from diverse sensors. The goal is to provide Original Equipment Manufacturers (OEMs) with a flexible foundation to simplify the integration of sensing, processing, and action in real-world systems. They plan to create modular reference systems accessible through Intel's Open Edge Platform and Altera SDKs. This partnership aims to accelerate the transition of physical AI systems from prototype to market deployment by addressing performance, power, and cost challenges.
Axon invested $1.2 billion in Skydio, gaining 20% ownership of the drone manufacturer. Axon will become Skydio's exclusive distributor for public safety in the U.S. and Canada to deploy Skydio drones to public safety agencies. This partnership aims to expand Axon's Drone as First Responder (DFR) program by leveraging Skydio's AI-backed autonomous technology. Skydio drones have reportedly reduced public safety response times by up to 65% in some deployments. The transaction is expected to close in late 2026 or early 2027.
HD Hyundai Robotics developed a robot with its own in-house foundation model, featuring dual-arm hands with five-fingered grippers. This robot will initially operate on a wheeled base for kitting before expanding to welding and painting. The company's AI agents have significantly reduced problem resolution time, cutting it from 21.5 days to 7.9 days. Skild AI is seeking partnerships with Korean motor and actuator makers to supply the hardware for its robot models. Skild AI pitches Korea as a strong source for robot hardware, aiming to integrate with Korean manufacturing capabilities.
Gritt and MasTec have partnered to deploy Gritt's AI systems on MasTec Renewables solar construction projects. Gritt develops general-purpose construction intelligence that uses robotic attachments to improve the safety, efficiency, and productivity of construction crews. The AI learns physical tasks like picking, placing, and assembly, guided by craft employees on job sites. These systems have successfully operated in diverse environmental conditions and monitor site activity to aid planning. The collaboration aims to apply construction intelligence to accelerate infrastructure building through future projects.
Oversonic Robotics is expanding its international presence by showcasing its RoBee R20 humanoid robot at electronica in Munich and Web Summit in Lisbon. The company specializes in designing and manufacturing cognitive humanoid robots in Italy, developed on a platform combining AI and human-robot collaboration. RoBee R20 is designed to assist workers in tasks like picking and handling in industrial environments, operating reliably in ISO Class 6 controlled settings. The company aims to leverage this experience to position Italian robotics solutions globally. This international exposure supports Oversonic's goal of demonstrating Italian expertise in the rapidly evolving humanoid robotics market.
ASICS is using Corvus Robotics autonomous inventory drones inside its distribution center to improve inventory management. These drones increased counting frequency sixfold in high-risk storage areas, leading to a 60 percent faster recovery of misplaced inventory. This process redirected approximately 200 hours of monthly counting labor toward inventory research and exception resolution. The system has successfully surfaced misplaced cases and improved inventory accuracy, which positively impacts revenue by enabling full-price sales. The drones use computer vision to count items in narrow aisles, reducing manual effort and improving fulfillment timelines.
AGIBOT proposes Worken, a framework to measure the useful work robots actually complete, moving beyond simple task completion. Worken calculates productivity using the formula: Active Robots $\times$ Autonomous Working Hours $\times$ Output Efficiency Ratio. This framework accounts for how independently robots operate and the quality of the output against set benchmarks. It aims to evaluate the practical contribution of physical AI in real-world settings by focusing on effective work delivered. Worken seeks to provide a clearer measure of robotic productivity, acknowledging that operating time alone is not equivalent to productive output.
Greenzie technology has achieved 255,000 cumulative autonomous mowing miles and covered over 9 billion square feet. This achievement stems from commercial deployments across more than 200 customers in 45 US states and four countries. The platform enables commercial mowing equipment to operate autonomously while supporting crews and increasing productivity. Greenzie partners with OEMs like Wright Manufacturing, Mean Green Electric Mowers, Greenworks Commercial, and Bobcat to integrate autonomy into equipment. The company also provides an adoption program to help customers successfully implement autonomous mowing solutions.
Ring is launching a new robotics division with a home security drone and an enterprise security robot. The home security drone, Ring Drone Cam, can fly between rooms to inspect the house and alert the user, integrating with other Ring devices. The enterprise robot, Ring Rover Cam, can patrol business routes, stream 4K video, and communicate via two-way audio. Both devices utilize a proprietary AI system called AI Pilot to autonomously map their environments. Ring asserts that the recorded videos are protected by TAKE encryption for privacy.
Jens Peder Kristensen cited NVIDIA results regarding the GR00T 2 Turbo's performance on a Unitree G1 humanoid. The model successfully completed 48 popcorn serving trials with a 93% success rate in clearing tables of objects. For onboard operation, the prediction time for the GR00T 2 Turbo running on NVIDIA Jetson Thor was 124 milliseconds. The full model achieved a 95% success rate across 50 manipulation tasks after fine-tuning. Kristensen highlighted the model's potential for skill transfer learned from human demonstrations.
GO 3 is AgiBot's newest embodied AI foundation model, succeeding GO 1 and GO 2 in the Genie Operator series. It was developed by the Chinese humanoid robot manufacturer and is based on the WRAM framework. The model achieved a Humanoid Guide Brain Score of 6 out of 10, indicating a capable level. Evaluation of such models should separate the software from physical robotic systems and consider real-world operational requirements. AgiBot plans to release GO 3 as open source by the end of 2026.
π*0.6 is a robot foundation model from Physical Intelligence, introduced on November 17, 2025, operating at the software layer of robotics. It develops using Recap, a training method that emphasizes continued training through human demonstrations, corrections, and practice. The model currently has a Humanoid.Guide Brain Score of 4 out of 10, indicating it is in the developing range. Its significance lies in its training methodology within the field of robot foundation models. Prospective users should contact the provider for details regarding access and deployment.
Google DeepMind released Gemini Robotics 2.5 and partnered with Apptronik and Mercedes-Benz to deploy Apollo 2 humanoids on production lines in Hungary and Germany. These robots autonomously perform automotive kitting by picking and sequencing metal brackets from bulk bins with a 96% success rate. Post-training improvements were substantial, with little performance loss when transferred to Mercedes-Benz production. Gemini Robotics 2.5 supports various robot configurations, including Apollo 2 and Franka Robotics Duo. The system also incorporates reasoning guidance to avoid hazards and supports model development through simulated practice runs.
SimEX is a simulation-based auto-research framework designed to help coding agents acquire physical capabilities for controlling real robots. It operates in two stages: first, the agent performs open-ended probe-and-optimize iterations in simulation to develop a robot skill toolbox. Second, the agent adapts this toolbox by performing a few physical trials, using simulation to correct the simulator and diagnose failures. This method allows agents to efficiently learn complex skills for tasks like towel folding and plate manipulation with minimal real-world interaction. The research suggests that simulation is a critical component for achieving physical intelligence by enabling agents to develop knowledge and procedures.
Maze Runner presents an end-to-end autonomy stack for navigating mazes using line-based localization and control barrier functions. Perception extracts wall lines from LiDAR scans using RANSAC and total least squares to estimate per-line uncertainty. Two Extended Kalman Filters are used for localization, one tracking a line map and another performing online SLAM. Planning uses an inflated occupancy grid with A* or Hybrid A* for path generation, while a Nonlinear Model Predictive Controller tracks the path with CBFs ensuring safety. Simulation results show high accuracy and fast planning/control execution times.
CAFE addresses scene-driven robot autonomy by formulating and executing tasks without instance-specific instructions. It maintains an online memory of objects, functional zones, and execution states to guide the process. A rule-guided Trigger-LLM facilitates task formulation from accumulated evidence, with Draft, Complete, and Audit modules proposing goals and checking dependencies. Outcomes update the memory, enabling correction and subsequent discovery. Evaluations on 12 scenes show that the Tri-LLM significantly outperformed baselines in task complexity and execution metrics.
Robo-MLT is a dual-system framework designed for reliable, long-horizon robotic manipulation in medical laboratory tasks. It combines explicit workflow monitoring with learned action generation to handle dynamic changes during execution. A vision-language planner generates skill sequences from protocols and observations, while a monitor tracks skill completion and triggers replanning. A skill-conditioned policy generates manipulation trajectories coordinated by an asynchronous runtime. Evaluation showed Robo-MLT achieved high task success rates by leveraging explicit workflow monitoring for enhanced reliability.
Learning-based collision avoidance methods must handle state uncertainty and heterogeneous agent dynamics reliably. The Dynamic Oriented Bounding Box (DOBB) is introduced as an online geometric representation that incorporates state estimates, uncertainty, and agent characteristics into a dynamic safety margin. DOBB derives geometric features characterizing agent interactions and a reward function that maintains physical meaning. Evaluation across four environments showed DOBB achieved a low mean total failure rate of 2.6% under out-of-distribution shifts. This robustness stems from the geometric abstraction of the safety margin, reducing coupling to specific uncertainty or actuation regimes.
Vision-language-action (VLA) models achieve high performance, but the extent to which this relies on training demonstrations is unclear. Retrieving nearby demonstrations and reusing their action chunks achieves 91.7% success on the LIBERO benchmark without policy training. Fine-tuning often improves the vision-language backbone while keeping actions close to demonstrations across various VLAs. A metric called Local Action Consistency (LAC) was introduced to measure action similarity between nearby observations. LAC tracks retrieval success and can guide the training of a retrieval key.
Ground Reaction Force-Aware Model Predictive Control (GRF-MPC) is presented as a framework for upright physical assistance by incorporating live ground reaction forces and center of pressure into a human dynamics model. This method formulates a real-time quadratic program to optimize assistive forces over a receding horizon. Experiments on an underactuated robot demonstrated that GRF-MPC improved tracking, reduced mechanical work, and increased voluntary excursion compared to a fully actuated baseline. The control system achieved fast solving times, enabling high-frequency replanning while providing stronger assistance with fewer actuators. Overall, integrating human-ground interaction results in more transparent physical assistance.
Latent Barrier Steering (LBS) is a hierarchical framework for safe flow-matching-based generative planning. It decouples behavior selection from path certification by steering the generator's latent toward larger safety margins first. Subsequently, it applies a path-space Quadratic Program only for residual safety violations. This approach increases the margin for the path corrector, leading to finite-time recovery guarantees. Experiments demonstrate that LBS improves goal reaching and reduces correction burden compared to traditional path-level safety filters.
Pretrained world models can guide next-best-view selection for efficient robotic 3D reconstruction in active sensing. A deployment adaptation methodology is introduced that learns a sensor-return adapter using limited local calibration data. This adapter estimates measurement dropout across predicted surfaces, enabling the selection method to favor views that fill data gaps. Experiments show the adapted method achieves higher surface coverage with 60% fewer captures than nominal selection. This demonstrates that robots can effectively reuse predictive models while adapting decisions to local operating conditions.