Teradyne Robotics A/S has filed a patent infringement case at the Unified Patent Court against a German subsidiary of the Chinese robot company JAKA. The lawsuit concerns alleged infringement of patents, covering both software and hardware related to Universal Robots. Teradyne Robotics asserts its commitment to enforcing intellectual property rights against unlawful copying of protected technologies. This marks the second IP infringement case in Europe for Teradyne Robotics in 2026, following a similar action in Germany. A ruling in this case could have effect across many EU Member States.
A Dockerized bridge is presented to facilitate the gradual migration of legacy ROS 1 Noetic systems to ROS 2 Jazzy on Ubuntu 24.04. This solution addresses the business challenge of avoiding immediate, costly full codebase rewrites. The bridge compiles ROS 1 Noetic from source alongside ROS 2 Jazzy within a single container environment. It provides a fully configured `ros1_bridge` with custom message mapping and includes E2E integration tests. The project is designed for optimized build times and bidirectional communication testing.
COMPASS adapts the pretrained NVIDIA X-Mobility policy into residual specialists for specific robots and environments using reinforcement learning. An agent-driven workflow automates environment validation, testing, training, and evaluation with human approval gates. This framework leverages a reference robot, like Boston Dynamics Spot, across various scenes, including generated or captured environments. The process involves training a residual policy, evaluating it against the base policy, and finally deploying it to the robot at runtime. Navigation capability is achieved by integrating perception and motion to allow robots to localize, route, and avoid obstacles safely.
Hyundai's CEO stated that robots are not replacing humans, a point made one day after a strike demanding job protections against automation was settled. The company plans to build a large-scale commercial robotics manufacturing facility in the US, with production starting in 2028. The initial 25,000 committed units will be primarily used by Hyundai's own factories, accounting for 83% of first-year output. The union agreement regarding automation only commits the parties to discuss employment when new businesses are rolled out, rather than guaranteeing job protection. Overall, the article details Hyundai's robotics roadmap and the limited employment context surrounding its automation plans.
Agibot debuted at the 2nd World Humanoid Robot Games, winning a total of 46 medals, including 18 gold, 16 silver, and 12 bronze. The competition tested humanoid robots in areas like sports, manipulation, and real-world tasks. Agibot's successes were in dexterous manipulation, scenario-based tasks, and motion-control events. The participating robots included models already in mass production, such as OmniHand, G2, A3, and X2. The results highlight Agibot's focus on applying embodied AI across diverse real-world scenarios.
ROS2 Connect proposes a WebSocket-based framework to enable secure and transparent ROS2 interaction across wide-area networks (WANs). It overcomes limitations of standard DDS multicast discovery mechanisms in remote environments. The system uses a client-server architecture to allow bidirectional exchange of topics, services, actions, and system data with integrated authentication. Experimental results show ROS2 Connect offers lower latency, greater stability, and improved scalability compared to existing solutions. This work establishes a reliable foundation for teleoperation and distributed robotics over WANs.
The paper proposes a method to extend ground-constraint LiDAR-IMU calibration to environments with tilted surfaces using a continuous-time framework. It addresses the limitation of existing methods that assume a flat ground by developing ground-plane residuals applicable to planar motion on tilted surfaces. The proposed method avoids the assumption of collinearity between gravity and the surface normal vectors. Experimental results show improved repeatability for both tilted and flat-ground scenarios, with significant gains in the tilted case. The method was validated using data from a Husky vehicle and offroad vehicles.
Autonomous driving planning often suffers from domain shifts when moving to new environments due to limited local data collection. SkyDrive introduces a framework using drone-based traffic monitoring to provide scalable supervision for autonomous driving agents in new cities. Aerial platforms naturally observe multiple road users simultaneously, creating a scalable source of grounded driving behavior. The research extracted 650K driving samples from 137 hours of footage to build benchmarks for trajectory planners. This work demonstrates that aerial traffic monitoring is an efficient method for adapting autonomous driving systems to new cities.
Sequential Object Placement Optimization with Convex Decomposition introduces a framework to optimize robotic object packing by framing placement as a differentiable nonlinear optimization problem in a decomposed free space. The method leverages the property that placing a convex object within a convex hull constrains its vertices, allowing for closed-form constraint calculations. A custom solver within this framework achieves optimal placement in milliseconds, offering a $100\times$ speedup over classical grid search. The framework is generalized to problems like 2D Tangram, 2D Tetris, and 3D Bin Packing, demonstrating strong packing utility. Experimental results include solving a real-world Tangram puzzle using robotic arms.
A control-oriented framework is presented for dynamic tracking and stability analysis of soft pneumatic actuators due to their nonlinear dynamics. This method models the actuator behavior using a nonlinear static equilibrium model and a linear residual dynamics model identified via Extended Dynamic Mode Decomposition with control (EDMDc). This representation allows for feedforward compensation, task-space feedback control, and local closed-loop stability analysis. Experiments showed tracking errors of approximately 1 mm RMSE at low speeds and below 10 mm RMSE at higher speeds. The approach successfully handles dynamic references and obstacle avoidance while predicting stable, marginal, and unstable operating regimes.
CRESSim-Neo is a batched GPU simulation engine designed for surgical robotics and robot learning applications. It integrates position-based simulation of rigid bodies, deformable tissues, fluids, and strands with batched rendering and a GPU-resident data pipeline. The engine supports complex applications like tissue manipulation, fluid suction, and ultrasound image synthesis. It enables GPU-resident robot learning through direct access to physics and rendering buffers, supporting zero-copy PyTorch integration. The system demonstrates high performance in various simulation tasks, providing a unified platform for surgical simulation and synthetic data generation.
Learning from demonstration (LfD) methods are explored for teaching robots skills without explicit programming, often evaluated in controlled lab settings. This research investigates the barriers non-expert care providers face when teaching robots in a home environment without live expert feedback. The study involves multiple visits with care providers, supported by pre-training and adaptive feedback guidance. The research aims to improve the ecological validity of LfD evaluations in a natural setting. A dataset from this longitudinal learning process is proposed for open-sourcing.
A voice-controlled, tendon-driven bionic hand was designed for coordinated finger movement using an efficient actuation mechanism. Servo motors generate finger movement guided by calibrated angular control, managed by a microcontroller programmed with Arduino. The system features a voice command interface to execute programmed gestures like open hand, fist, pinch, and half flexion via a Bluetooth architecture. Experimental testing showed the system achieved repetitive, precise motion with stable grasping of various objects. These results demonstrate a simple, efficient mechanical and control architecture for predictable hand control.
A generative action-chunk sampling framework is proposed for adaptive stiffness control in physical human-robot collaboration. The system conditions on RGB images and joint-torque estimates to sample future action chunks from a learned prior. Variation among these sampled actions is used to continuously adjust joint stiffness and damping. Increased variation results in greater robot compliance to better facilitate human guidance. This method demonstrated superior success rates in a collaborative transport task compared to fixed or deterministic baselines.