Native ROS 2 is being implemented on the Zephyr operating system using Cyclone DDS as the underlying middleware. The current software stack involves rclc, rcl, rmw_cyclonedds_c, and Cyclone DDS running on an ESP32-S3 via Wi-Fi. This setup allows the board to join the DDS/RTPS network directly without a Micro XRCE-DDS Agent. Testing has been performed with the stock ROS 2 Lyrical stack, supporting fixed-size messages and best-effort QoS. Future work involves extending this path with features like Reliable QoS and DDS Security.
A demonstration utilizes the new SMACC PX4 Client Library to autonomously control a drone within the Gazebo simulation environment. This involves simulating the SMACC and PX4 flight stack specifically in the Strait of Hormuz. The source code for this application is available on GitHub. The process and usage instructions are detailed in the application's readme file. This work connects ROS2, PX4, SMACC2, and Gazebo simulation.
AthenaZero is a bimanual robot designed to minimize inertia while maintaining control authority for dynamic manipulation. It achieves an endpoint mass significantly less than conventional robots by using quasi-direct drive actuation and transmission remotization. This design provides inherent torque transparency, making it ideal for dynamic tasks. Demonstrations on baseball-inspired tasks showed speeds exceeding 30 m/s for throwing and 14 m/s for catching and batting over short distances. The system effectively handles high acceleration required for complex interactions in human-comparable timeframes.
The DITTO framework addresses the challenge of collecting data for high-DOF hand manipulation by creating a dexterous interface for transparent teleoperation. It achieves this through the co-design of a 7-DOF robotic hand and a kinematically equivalent motorized exoskeleton. A 1-to-1 actuator mapping allows for both handheld data collection and bilateral teleoperation with unified, joint-level force feedback. This system spans the operator's natural index-to-thumb workspace. Learned policies are demonstrated to showcase the DITTO system's dexterous capabilities on contact-rich tasks.
ULOHA is an underwater bimanual robot learning platform designed to address the gap in underwater visuomotor policy learning compared to single-manipulator studies. It integrates custom hardware, software extensions to LeRobot, teleoperation, multi-view sensing, and demonstration collection. Experiments demonstrate coordinated underwater bimanual behaviors such as inter-arm transfer and shared-object manipulation. The system evaluates learning methods like ACT and Diffusion Policy while investigating the effects of the underwater environment on execution. ULOHA provides a unified platform for studying underwater bimanual robot learning under coupled perceptual and physical effects.
REACT is a fully spiking state-space model designed for real-time, event-driven temporal perception in robotic systems. It processes raw events individually without temporal accumulation, addressing the integration delay found in frame-based methods. The model utilizes a complex-valued spiking neuron, C-SiLIF, whose dynamics are governed by the physical inter-event interval. Evaluation on gesture recognition and time-to-collision estimation shows REACT achieves low-latency, accurate perception. These results demonstrate that event-driven spiking dynamics enable low-latency, continuously updated temporal perception.
Research on 4D millimeter-wave radar perception algorithms spans signal processing to dynamic scene reconstruction for autonomous driving. The review organizes the field by perception tasks, covering object detection, motion estimation, spatial perception, and dynamic scene understanding. It compares various approaches including radar-only learning, multimodal fusion, and cross-modal supervision. The review emphasizes exploiting elevation, Doppler, and radar physical priors across these tasks. Finally, it summarizes datasets, challenges, and future directions for 4D radar perception.
The research addresses grasping objects with uncertain shapes, sizes, and poses by focusing on robust closing motions. A motion planner was designed based on coarse object templates, human grasp types, object-centric interaction, and compliant, sliding contacts. Evaluation showed the planned motions tolerated significant size and pose errors without feedback, exceeding a data-driven planner in many conditions. The method successfully grasped 82.5% of everyday objects autonomously. This work suggests incorporating robustness into closing motions to improve reliable manipulation.
Standard reinforcement learning navigation pipelines struggle with complex humanoid terrain like stairs due to difficult interactions. The research decomposes the navigation problem into training a policy on reduced-order dynamics using full 3D LiDAR and then transferring this knowledge to the full-order humanoid dynamics. A Poisson safety filter applied to the navigation policy output successfully maintains safety against out-of-distribution obstacles. This approach was demonstrated on a Unitree G1, enabling mapless multi-floor navigation with significant vertical and path displacements. The method successfully learns safe navigation by leveraging simplified models to manage complex physical interactions.
OHRID Retail is an open, multimodal dataset documenting human activity in retail environments involving locomotion, reaching, and object handling with mobile robots. It was collected from 16 healthy adults performing a simulated shelf-picking task under three conditions of robot guidance: none, low speed, and high speed. Data was collected using Xsens inertial sensors for whole-body kinematics and Delsys surface electromyography sensors for muscle activity. Descriptive analysis revealed variations in movement intensity and muscle activation based on the robot interaction conditions. The dataset supports research in areas like human activity recognition, multimodal fusion, and human-robot interaction in retail settings.
A morphing aerial robot is proposed using thruster-integrated flexible continuum links for shape-adaptive aerial manipulation. These flexible links act as soft manipulators, passively conforming to the environment's shape. Thrusters distributed along the links expand the feasible wrench space, allowing for larger interaction forces. A composite leaf-spring structure and control methods are used to suppress vibrations and ensure stable flight. This work presents the first realization of aerial manipulation using flexible links with integrated thrusters.