Big drones like Anduril's Thunder are becoming increasingly common, with features that surpass traditional military hardware. Autonomous drone-boat company Saronic has also had recent successes. Meanwhile, companies like Ramp are developing AI models to improve efficiency and reduce costs, such as an internal AI model router that claims to drive a 30% reduction in AI spend.
Generative Bionics unveiled the Gene.01 humanoid robot with full-body tactile sensing at AMD Advancing AI 2026. Built in six months, the robot integrates physics-native AI, body, mechanics, and intelligence to better interpret and respond to the physical world. The platform is scalable, customizable, and open-sourced, enabling collaboration with people over isolation and addressing problems video-only learning cannot solve. Gene.01 will serve as the base for a family of use-case-specific humanoid robots, with plans to adapt the platform for various industrial applications and build a sovereign European physical AI ecosystem.
Current world model conditioning in robotics relies primarily on visual observations from cameras and end-effector position from joint encoders. However, this is not sufficient for contact-rich manipulation tasks, as touch information is invisible to cameras and unresolvable from joint position alone. The missing quantity in these models is μ, the coefficient of static friction, which determines whether a grasped object stays grasped and whether the grip force is sufficient given surface properties, load, and contact geometry. Contactile's PapillArray Tactile Compute Module fills this gap by computing the full set of physical quantities required for complete contact conditioning.
Current limitations of the community event calendar include not being able to select only desired events and receive updates for those selected events. The "Copy to my calendar" feature does not sync with the original event, even if it's recurring. Adding the entire calendar to Google Calendar also includes all events, rather than just the user's interests.
Nvidia Jetson kits and modules have experienced a 50-100% price increase across their lineup. Alternative boards and SoCs used for running ROS in real-world robots include the AMD Strix, Intel robotics computers, and upcoming Qualcomm Arduino hardware, catering to both students and larger fleet deployments. These alternatives aim to provide an affordable and viable option for those unable to afford the increased Jetson prices.
Siléane, a France-based robotics specialist, partnered with Hottinger Brüel & Kjær (HBK) to deliver an automated gripping solution for a global cosmetics manufacturer. The project embeds smart sensing into the end-of-arm tooling, providing real-time grip quality validation and ensuring precision and reliability in high-speed packaging lines. By integrating force measurement directly at the point of interaction, the system captures real-time data on physical contact, reducing product loss and tray deformation.
Industrial robots paired with laser cleaning systems optimize production lines by automatically removing burrs from metal components, saving time and effort. Laser cleaning uses a powerful beam to remove unwanted materials from metal surfaces without causing damage. The combination of robots and laser cleaning saves human effort, improves safety, protects metal surfaces, and enhances welding quality.
A nested kino-dynamic framework is proposed to rapidly check feasibility and generate dynamically consistent trajectories given a candidate contact sequence. The framework integrates a feasibility-guided tree search with a Large Language Model-based contact plan sampling strategy. It has been demonstrated that the resulting trajectories can be tracked using a reinforcement learning-based controller, achieving high-quality results for execution in real-world loco-manipulation scenarios.
Reinforcement learning (RL) for legged robots is advancing locomotion by adapting to new terrain. The RL framework explores torque control for heavyweight high-torque quadrupeds, achieving speeds of 3.5 m/s and 1.5 rad/s on the Unitree B1 quadruped. The quadruped can also walk up and down stairs without an exteroceptive sensor.
A physical platform called the Open Ant has been developed to simplify the transition from reinforcement learning research in simulations to real-world applications. The platform features a robot designed for walking and supports various RL algorithms, including SARSA(λ) and Soft Actor-Critic (SAC). Competent walking policies can be learned by the robot in approximately one hour directly from its experience with the physical platform or simulation. Both hardware design and software are available as open-source on GitHub for customization.
Researchers have formulated target interception in crowded environments as a partially observable Markov decision process and solved it online using tree search under a fixed computational budget. The approach allows robots to reason over multiple possible future interaction outcomes with human agents, which is crucial since their navigation intent is not directly observable. A unified planner that jointly branches over steering and speed performs better than a sequential path-speed planner at higher crowd densities, revealing the structural limitation of spatial restriction in the latter.
A novel robot, DASH (Ducted Aerial Spring Hopper), is presented with a minimalistic design that enables both aerial and ground locomotion capabilities. The robot features a ducted fan co-axial body with a springy leg at the bottom, providing thrust-vectoring for agile flying and efficient ground hopping. A contact-implicit model predictive controller automatically chooses locomotion modes and actuation, achieving optimal locomotion in two modes, including periodic hopping, aerial flight, and mode-free locomotion with autonomous obstacle traversal.
A line-scanning lidar yields range and azimuth values in a fixed plane, requiring relative motion between the lidar plane and the object to perceive it in 3D. To achieve this, a rotary platform setup can be used, but accurately identifying the axis of rotation with respect to the lidar coordinate system is crucial for precise reconstruction. This work explores automated approaches to statically and dynamically estimate the transformation of a rotary platform's axis of rotation with respect to a static line-scanning lidar, using real-world datasets from a custom made rotary platform and an FMCW lidar. The proposed algorithms have been validated on these datasets and their convergence characteristics studied for various initial conditions.
STeP proposes a framework using Signal Temporal Logic (STL) to connect high-level language understanding with low-level robot execution. The approach leverages a Vision Language Model (VLM) to decompose language instructions into subtasks, generate STL specifications, and select a low-level policy for each subtask. STL specifications translate language-derived intent into precise constraints, which are enforced directly or monitored during execution using a common formal structure. This enables precise specifications for action generation with vision language models in real-world tabletop domain.
Researchers Hossein Maghsoumi and Yaser P. Fallah have framed sim-to-real transfer in autonomous racing as a full-stack, real-time systems problem. They introduce a structured three-layer perspective to analyze how mismatches propagate and amplify through closed-loop feedback, presenting diagnostic metrics beyond nominal lap time, including performance flip, stability-oriented measures, sensitivity to delay and noise, and latency distribution characterization. The resulting framework clarifies cross-layer failure mechanisms and provides practical design principles for deployable autonomous racing systems operating near dynamic limits.