A drone called Ice Dart has been designed with spiny feet that allow it to latch onto glaciers and drifting icebergs, enabling long-term monitoring of these areas. The drone features shock-absorbing legs and claws, which enable it to perch on steep and slippery icebergs.
Unitree Robotics, a Chinese hardware manufacturer, made its highly anticipated IPO on the Shanghai STAR Market, with its shares soaring by up to 629%. The company is highly vertically integrated, producing 86% of its bill of materials in-house, and has secured strategic partnerships to solve industry software bottlenecks. Unitree's rapid growth contrasts with recent US import bans targeting Chinese robotics, underscoring a deepening global divide.
Foxglove has introduced several new features, including a search experience that allows querying recordings with combined topic and semantic search, an A/B comparison feature for visualizing and comparing robot runs, an agent sidebar that enables tasks such as creating layouts and triaging issues, and improvements to remote access and Bring Your Own Server (BYOS) functionality, including optimized WebRTC streams and the ability to use a single storage bucket.
Researchers at OSRA, in collaboration with the Special Interest Group on Physical AI, have published a blog post detailing their efforts to integrate Physical AI with ROS. The SIG PAI explores the future of Physical AI and ROS, proposing that leveraging the strengths of both is the most effective path forward. The group shares their findings, ongoing work, and long-term roadmap for achieving this integration.
Faraday Future Intelligent Electric (FFAI) has announced its support for AIxC Holdings' orderly exit from its digital asset treasury strategy, focusing on physical AI and robot sharing and rental. RoboShare, AIxC's robot-sharing and rental platform, completed its first paid commercial order, marking the platform's transition into commercial operations and initial revenue generation. The platform will provide customized performance production and content design services for its first customer, Los Angeles-based artist DU$TY, and expand its product offerings and develop the market. FFAI will provide AIxC with strategic, product, technology, business, and capital-market support to help the company establish a sustainable business model centered on transaction and service revenue.
Hexagon Robotics and Schaeffler have announced the next step in their collaboration, with AEON humanoid robots training at Schaeffler's Humanoid Gym in Germany. The gym provides a dedicated industrial environment for training and validation, enabling AEON to learn new industrial tasks through imitation learning and repeated execution. Schaeffler will supply advanced actuator technology to AEON, supporting its industrial deployment across its manufacturing operations. The training program aims to accelerate AEON deployment across multiple use cases in manufacturing workflows within the next six months.
Scientists have developed a hybrid robotic skin combining electrical impedance tomography and pneumatic pressure sensing to improve force detection across large areas of a humanoid robot. The system uses EIT to determine where contact occurs, while four sealed pneumatic pads provide an estimate of the force being applied. The researchers report a reduction in sensitivity variation, with the coefficient of variation falling from 0.31 to 0.14. The approach takes advantage of the different strengths of the two technologies, allowing EIT to provide spatial information while pneumatic sensing provides a net-force measurement.
Autonomous robots are being tested on construction sites to navigate around moving crews without relying on GPS. Simultaneous Localisation and Mapping (SLAM) technology, which uses LiDAR to build a map in real-time, is being used to solve the "where am I" and "what does this space look like" problems. The technology has been deployed at a construction site in Oxford and has mapped the site and carried out autonomous inspection within four hours. However, the technology has limitations, including issues with airborne dust and physical debris, and deployment cost and integration with existing site management software.
Rodriguez has introduced a collaborative linear transfer system, the Movotrak Cobot Transfer Unit (CTU), to allow a cobot to move between multiple workstations and extend its operating range to up to 10 meters. The system includes integrated collision detection, enabling the carriage to automatically stop in case of a collision while moving. This allows a single cobot to perform tasks at different positions along a production line, increasing robot utilisation and reducing changeover requirements. The system is designed to be compatible with standard cobots and can be installed horizontally or mounted on a wall or ceiling.
Brightpick, a provider of AI-powered warehouse automation, is expanding its partnership with Dr. Max Group, Europe's largest pharmacy chain, by deploying 44 additional autonomous mobile robots across facilities in the Czech Republic and Slovakia. The total number of deployed robots will reach 74 units. The partnership has led to a 95% reduction in manual picking labor and a 40% increase in warehouse capacity. The expanded partnership includes new projects in the Czech Republic and Slovakia.
Vrije Universiteit Brussel (VUB) leads a four-year European research project called RESSORT, which aims to make soft robots more sustainable, intelligent, and reliable for applications in healthcare, the food industry, and industrial inspection. The project combines breakthrough technologies, including self-healing materials, variable-stiffness structures, and advanced multi-material 3D printing techniques. Researchers will develop demonstrators, such as a personalized soft exosuit for healthcare, an adaptive robotic system for the food industry, and a deformable drone for inspection tasks in confined environments. The project is funded by the European Union's Horizon Europe research and innovation programme and brings together 14 partners from Belgium, Denmark, Germany, Greece, Norway, Sweden, and Switzerland.
Machines that kill without distinction, answerable to no one, leave no one safe. Autonomous drones, loaded with AI, pose a significant risk to civilians and leave no one safe. The principle of distinction, which requires a gap between a target and a lawful target, is difficult to close with autonomous drones. If we cannot teach a drone to keep the laws of war, we will not teach a greater intelligence to keep faith with all we value.
Galbot, a Beijing-based robotics startup, has previewed its fully bipedal humanoid robot, the ET1, ahead of its formal debut at WRC 2026 in Beijing. The ET1 showcases dynamic athletic capabilities through a CGI teaser, highlighting its ability to swing a tennis racket. The robot is based on the LATENT athletic framework, which was introduced earlier this year and validated on a third-party platform. Galbot's expansion into bipedal locomotion reflects the intensifying competitive pressure in the humanoid sector, with the company eyeing a potential $4 billion IPO.
A3 and Mahoning County Career & Technical Center will offer no-cost robotics and automation training to Ohio residents through a $499,000 workforce grant. The training will cover industry-recognized topics and serve students, educators, adult learners, and manufacturing employers. The program aims to help small and medium-sized manufacturers develop a skilled workforce to adopt robotics and automation technologies effectively. The training will be available online and can be signed up for at https://www.automate.org/training/ohio-training.
Researchers Omar Rayyan and colleagues developed FetchMan, a visual humanoid loco-manipulation policy that learns from simulated experiences and can generalize to novel scenes and objects. They replicate a "recipe" for learning from simulated data, first using behavioral cloning, then breaking through a performance ceiling with reinforcement learning. Refining the cloned policy with Flow-GRPO resulted in performance that synthetic behavior cloning couldn't match. FetchMan was evaluated on a simulation benchmark, deployed on a real humanoid robot, and extended to multi-object training.
A proposed controller, $\lambda$-hold control, enables a muscle-actuated skeletal model to learn human-like sprinting using only a minimal reward within an hour of training. The policy uses a per-muscle equilibrium-point threshold length $\lambda$ and a stretch-reflex recruitment law to compute muscle excitations automatically. This approach, grounded in physiology, sharpens the exploration through the proposed $\lambda$-hold controller, bringing together the equilibrium-point hypothesis, intermittent control, and optimal feedback control. The controller enables human-like movement in predictive musculoskeletal simulation.
This paper proposes a hybrid framework that integrates low-resolution Digital Elevation Model (DEM) data with real-time LiDAR-based obstacle detection and terrain analysis for efficient path planning in autonomous UGV navigation. A global path is initially computed using a preprocessed DEM-based A* algorithm, and local sensor data drives adaptive path correction to maintain safety and efficiency. The integration enhances path efficiency and terrain traversability, achieving a 95% obstacle avoidance rate and reducing the average encountered slope from $8^\circ$ to $2.7^\circ$ in custom terrain.
PDDL-ART autonomously generates task-specific PDDL domain and problem descriptions from a single expert demonstration, natural language task description, and a library of high-level action names. The approach introduces a multi-stage correction pipeline operating at syntactic, semantic, and execution levels to ensure generated descriptions are semantically aligned. PDDL-ART achieves an average success rate of 93.3% on challenging manipulation tasks, compared to 78.3% for a baseline VLM-based planner.
Teach-and-Grow Learning (TGL) is an agent-centered architecture for general robot learning, presenting a multimodal agent that turns successful demonstrations into reusable Skill Blocks. These blocks are composed, selected, and revised based on physical outcomes, storing executable behavior and structured Experience Memory. The LIBERO evaluation attains state-of-the-art performance, exposing skill induction, persistent reuse, and agent-directed adaptation. TGL treats deployment as a period of continued learning, where one task can make the next easier.
Researchers developed a reinforcement learning-based method to enable brachiation on a life-sized dual-arm robot using Waypoint-Guided Reinforcement Learning (WGRL). The approach guides behavior acquisition by specifying waypoints for the end-effector trajectory while generating whole-body motion through reinforcement learning. The method achieves both forward progression and motion stability, and has been evaluated through Sim-to-Sim experiments and hardware experiments, confirming robust brachiation and failure recovery behavior.
ORPA, a framework for immediate, feedback-driven correction of robot actions, augments a pretrained control policy with a lightweight, feedback-conditioned module that predicts residual adjustments directly in joint space, allowing the system to adapt its behavior at runtime. This enables correcting robot actions without modifying the underlying policy parameters, addressing issues with traditional dataset aggregation and full-policy retraining. ORPA improves success rate and recovery from small perturbations compared to baseline control policies and rule-based inverse kinematics corrections. ORPA is evaluated on a set of precision-sensitive manipulation tasks using the ALOHA platform.
Researchers Jie Gu and colleagues introduce a new method for planning and reconfiguring motion primitives for deformable planar modular robots. They develop a square-cell abstraction that maps the robots' deformable geometry to a fixed-size grid, allowing for reversible and reconfigurable motion primitives. The new method proves that every non-straight edge-connected configuration with 7 or more modules can be transformed into a fixed canonical staircase, and provides a constructive planner that transports removable boundary modules while preserving connectivity. Experiments show the planner substantially reduces planning time.
MANIGUARD is a specification-grounded framework for evaluating and improving the safety of foundation-model manipulation policies. It comprises a benchmark task suite and a safety-annotated trajectory-generation pipeline, comprising 200 locked base tasks with safety specified independently of task success. The framework is tested on a Franka platform, with zero-shot and fine-tuned validation results showing that safety must be evaluated independently of task success, and fine-tuning raises safe task completion and engaged-and-safe behavior.