US Federal Communications Commission blocks foreign-made advanced robotic devices from receiving equipment authorization, citing supply-chain vulnerabilities and cybersecurity risks. China threatens to retaliate if US maintains the restrictions. Researchers develop a drone that blends into its environment, and a humanoid robot with a curved OLED "face" for communication. A new Bayesian model aims to help designers avoid uncanny valley issues in future robots.
Nucleus, a European startup co-founded by veterans from CERN, ESA, and other prominent robotics organizations, has successfully deployed humanoid robots into a real factory in under 90 days. The company is leveraging large-scale teleoperation and supervision to bypass the wait for full autonomy and generate training data from day one. Nucleus operates on a robotics-as-a-service model, billing by the hour for completed physical work, while building a proprietary dataset of long-tail industrial edge cases.
Researchers at MIT have developed an adaptive physical therapy system that uses generative AI to learn from physical therapists and interactively support stroke patients, teaching robots how to assist with physical and occupational therapy without replacing human care. The system combines transformer-based diffusion models with real-time force feedback, enabling the robot to safely adjust assistance based on each patient's capabilities. The approach has been tested in a clinical study and has potential applications beyond physical therapy, such as in industrial settings or collaborative workspaces.
Sarla Aviation, an Indian startup based in Bengaluru, is using Siemens Xcelerator to accelerate the development of its eVTOL aircraft. The company aims to launch operations by 2028 and build a comprehensive air mobility platform for urban transportation. By adopting a digital thread approach, Sarla Aviation can reduce design iterations, improve collaboration, and strengthen certification processes.
Intelligent automation combines machine learning, predictive models, and computer vision with traditional machinery to enable adaptive workflows. This marks a shift from rule-based automation toward self-optimizing systems capable of retrieval, reasoning, and autonomous decision-making. The adoption of intelligent automation is as much a change-management challenge as it is a technical one, requiring operators to trust the system's recommendations, maintenance teams to develop new skills, and leadership to communicate the shift clearly.
Avnet and Weston Robot partner to launch an edge AI inspection platform powered by quadruped robots. The platform uses Avnet's backpack with AMD Ryzen AI Embedded processors, delivering up to 50 TOPS of AI performance for real-time autonomous inspections. Powered by edge computing, the system can detect operational issues earlier, reduce costs, and improve operational continuity in GPS-denied environments.
The University of Florida has opened a robotics lab dedicated to industrialized construction, funded by a $1 million donation from Autodesk. The lab aims to address Florida's housing shortage and an impending surge in retirements among the construction workforce. A new degree program, launched this fall, combines advanced manufacturing principles with digital workflows and systems to educate students on precision, technology-driven, and sustainable building processes.
Available simulation assets are rarely usable as-is for production-oriented simulation. A robot gripper's behavior, such as contact behavior, is often incomplete or inaccurate in available resources like Isaac Sim and ROS ecosystems. The Robotiq 2F-85 gripper asset, tested against multiple candidates, exhibited limitations in capturing the gripper's physics.
A modular perception pipeline is presented for household manipulation tasks, focusing on dishware handling in kitchen environments. The pipeline integrates object detection, multi-view segmentation, 3D reconstruction, and feature fusion for 6D pose estimation and grasp planning. A custom kitchen dataset is used to evaluate the best-performing configuration, achieving an ADI of 89.12%. This configuration enables successful execution of tasks on physical robots without environment-specific retraining.
Researchers Nie Lin and colleagues developed SiMDex, a framework to select optimal human videos for training robot manipulation tasks by mining similar egocentric videos. The framework uses a three-layer pipeline to extract task-relevant subsets from a pool of 32M human samples, achieving an overall success rate improvement from 47.7% to 61.1%. SiMDex outperforms indiscriminate data mixing using only 1.49M mined samples (<5% of the pool).
Researchers developed and tested an embedded photoplethysmography (PPG) sensor in a pocket-sized robot to enable continuous heart rate monitoring during tactile interactions. The sensor was paired with an Inertial Measurement Unit (IMU) for motion artifact rejection, and a confidence-based smoothing algorithm for recovery periods. The filtering strategy significantly reduced Mean Absolute Percentage Error and achieved statistical equivalence to ground truth measurements. However, the sensor required stability over longer periods to converge, limiting its effectiveness in short-duration interactions.
SAFECAST improves failure detection in VLA policies using contrast-set training and calibration, outperforming a state-of-the-art baseline in DROID and LIBERO simulation experiments with significant RO-AUC scores. Contrast set perturbations are used to augment data, leading to better probes than real rollout data only.
Researchers have designed and flown a micro hovercraft propelled by an ion engine, leveraging ground proximity effects to increase efficiency. The vehicle, measuring approximately 1.6 grams, flew for extended periods while generating zero audible noise and withstanding numerous takeoff and landing cycles. Its thrust efficiency exceeds that of similarly sized electroaerodynamic robots by an order of magnitude, setting the stage for a new class of robot designs.
Omar Muhammetkulyyev, Oren Salzman, and Tichakorn Wongpiromsarn introduce epsilon-rule-dominance, a notion of approximate dominance under rulebooks, to address the computational challenge of computing full set of rulebook-optimal solutions. They propose RA*pex, a best-first search algorithm leveraging dimensionality reduction to efficiently compute compact sets of epsilon-approximate rulebook-optimal solutions.
SCOPE (Safety Certification through Observation Planning and Execution) is a path planning framework that decouples optimistic goal-directed guidance from certified execution in unknown 3D environments. It uses online safety-volume certification to ensure the complete robot-inflated volume of an intended motion can be observed and verified free before execution. The system converts uncertain points along optimistic routes into explicit observation obligations, resolves them through target-centric viewpoint search, and recursively clears intermediate obligations when useful viewpoints are not yet certified-reachable. SCOPE reaches 100% success rate in 60 randomized tasks while maintaining near-zero entry into non-certified inflated space.
SSC (Structured Subtask Chain) proposes a state-transition representation to bridge the extremes between natural language descriptions and rigid template formats. Each Structured Subtask Template (SST) stores core action components, flexible conditions, base-motion field, and after-state scene graph. SSC supports three vision-language assisted functions: rendering SSTs as natural language, checking against four state-transition rules, and completing underspecified fields through a query resolution cascade. The pipeline is instantiated on BEHAVIOR-1K for logic verification and content completion.