EdgeAI Forge is a local-first engineering architecture aimed at generating, testing, benchmarking, and documenting ROS 2 and machine-vision pipelines. The initial target workflow is a ROS 2 vision pipeline generator that produces reviewed artifacts, including package manifests, nodes, topics, parameters, tests, and documentation. The system includes safety boundaries to prevent automatic deployment of generated robot-motion or production-deployment artifacts without human review. Feedback from the ROS community is sought on various aspects of the project's design and implementation.
A 2nd-year Mechanical Engineering student at NIT Jalandhar is exploring Robotics and Automation, with goals of working on real robotics projects and pursuing research/internship opportunities. The student is looking to connect with students learning ROS 2, professionals working on interesting robotics projects, researchers/engineers open to collaboration, and those who have previously pursued robotics research internships. They are seeking beginner-friendly open-source robotics projects for contribution.
Researchers at UCL are seeking input from teams on how to detect degradation of perception or sensor-fusion stacks in production before it causes visible failure. They want to know what methods teams currently use, such as manual spot-checks, logging and alerts, or formal calibration schedules. The researcher is genuinely looking for knowledge and not promoting a product.
A US citizen with no US history is trying to break into the US robotics/embedded industry as a 4th-year Computer Engineering student from Palestine. The individual has relevant experience with ROS2, FreeRTOS, and a competition robot, but lacks local network and job experience due to the lack of an embedded or robotics industry in Palestine. They are considering looking for a job directly, trying to get into a US graduate program first, or another option.
LyEvO combines constrained Evolutionary Optimization and Statistical Model Checking with Lyapunov-based stability analysis to train safe and robust controllers for sim-to-real transfer. It leverages prior knowledge of system dynamics to compute an initial candidate stability region, which is iteratively expanded through operational scenarios and verification. This integrated procedure provides a practical criterion for assessing deployment readiness, demonstrated on Cartpole and 3D Quadrotor benchmarks.
A parallel motor-cable system with six motor-cable units is developed to provide external force without interfering with the movement of a surgical robot. This system includes controllable tension cables connected to the end-effector, allowing for accurate force actuation with errors less than 1 N in preliminary experiments. The development of the system involves motor-unit hardware, control software, sensor drivers, simulations, and more.
Researchers Varun Raveendra, Connor Mattson, and Daniel S. Brown introduced SyncSBC, a decentralized swarm behavior prediction system that combines machine learning and distributed consensus to classify collective swarm behavior and synchronize swarm decision-making in an entirely decentralized manner. The system achieves high classification accuracy and low synchronization delay, making it suitable for real-world deployment. It is also used to demonstrate two promising swarm applications on real robots, showcasing the ability of swarms utilizing SyncSBC to accurately identify anomalies in robot behavior and autonomously coordinate collective changes in swarm behavior.
A real-time Plan-and-Avoid (PAA) framework coordinates cooperative multi-agent airspace operations around a declared priority trajectory. The PAA framework predicts uncertainty-aware separation violations and generates vehicle-constrained unilateral advisories to modify nearby aircraft trajectories. The approach is tested using over 900 forced-landing cases, totaling over 140 hours of simulated flight, with a worst-case end-to-end response time of 5.7 seconds.
SoRoMoX, a Python/JAX framework, provides fully numerical, JIT-compilable models for soft robotics that support differentiable and parallelizable workflows. It implements articulated, Piecewise Constant Strain, and Variable Strain models through a unified interface, enabling faster rollouts on GPUs (up to 234.6x) and CPU (up to 18.1x). SoRoMoX improves performance in various soft-robot applications, such as static-equilibrium system identification and control-gain optimization.
CrossTracer, a hierarchical framework, adapts a pretrained VLA model to predict an initial navigation trace and refines it with embodiment-conditioned residual corrections. CrossTracer represents navigation plans as normalized image-plane waypoints, forming a unified pixel-space interface between semantic reasoning and physical grounding. It outperforms the strongest evaluated general-purpose baseline by 28.1% relative improvement in the NaviTrace benchmark, achieving a total score of 45.68.
A custom shroud system transforms a micro air vehicle into a multi-modal robot capable of operating as a high-efficiency hovercraft or a free-flying drone. The shroud system increases in-ground-effect force by nearly three times while maintaining comparable out-of-ground-effect aerodynamic thrust, and lightweight shrouds made from thin-film thermoformed components increase single-charge flight time by 60% in-ground-effect and only 30% in free-flight.
A new physically grounded JEPA world model, PSG-JEPA, outperforms state-of-the-art latent world-model baselines in three levels: latent identifiability via probing, goal-conditioned planning on frozen latents, and policy learning in simulation and on a real robot. The proposed model shapes its latent space with two complementary grounding objectives during training, improving downstream planning and policy performance.