X Square Robot completed a one-hour live logistics challenge, sorting 1,816 parcels with a success rate of over 98%. The dual-arm robotic system used purpose-built grippers instead of complex five-finger dexterous hands, prioritizing high-speed and cost-effective throughput. The company emphasized its commercial practicality in hardware choices, focusing on real-world deployment rather than marketing flash. This challenge followed similar formats by Figure AI and AGIBOT, showcasing the technical hurdle of using vision-based AI foundation models to process raw camera pixels and generate real-time motor torque for manipulation in unpredictable environments.
Join Waffle, a weekly triage meeting on Thursdays at 4:30 PM UTC via Google Meet, to contribute to ROS 2. The meeting helps maintainers quickly discuss and assign issues and pull requests that have gone unnoticed. No preparation is required, and participants can attend once or whenever they need to. It's an opportunity to see how the project works firsthand and find something to work on, especially if you're familiar with a specific corner of the stack.
Einride and DAF Trucks have partnered to integrate Einride's autonomous driving system, Einride Driver, into DAF's premium vehicle platform. The collaboration aims to accelerate the scale-up of SAE Level 4 autonomous electric freight transportation, addressing industry challenges such as driver shortages and rising operating costs. The integration will enable safe and scalable autonomous operations, focusing on repetitive tasks without human intervention. Initial tests are planned for 2026, with full integration and commissioning expected in 2027.
Honda's Avatar Robot and its human-scale multi-fingered hand are designed to be a physical extension of a human operator, allowing remote control over tasks that require human judgment combined with robotic reach and manipulation. The robot's hand combines dexterity, strength, and durability, capable of performing tasks such as threading a needle and handling heavy objects. Honda's research has led to the development of reliable hands for humanoid robots, which is crucial for making them commercially useful.
Humanoid robotics has reached a critical point where cognitive AI capabilities are outpacing hardware advancements. Traditional industrial robotic joints are insufficient for dynamic motion required in unpredictable environments. Quasi-Direct Drive (QDD) architectures provide low impedance, backdrivability, and precise force interaction for next-generation humanoid platforms. Integrated actuator solutions, such as those by CubeMars, combine motors, drivers, and encoders into compact systems to accelerate reliable real-world deployment.
Researchers Shreyas Kowshik and his team developed MiDAS, a simple offline-to-online RL recipe that adapts generalist robot policies with minimal data. The approach uses behavior cloning on single or few demonstrations to anchor a pre-trained policy to the target task, followed by value-based online RL improvements through autonomous interaction. With as little as one demonstration, MiDAS recovers strong task performance across various robot platforms and outperforms baselines, demonstrating reliable robot policy adaptation from minimal human guidance.
Researchers propose TrafficDiffuser, a top-down traffic scenario generation framework that jointly models initial and goal state pairs to generate high-level scenarios. The framework improves upon existing methods by making initial states better interpretable and reducing trajectory generation into an infilling problem. It outperforms the next-best approach for agent initialization, capturing real-world distributions on the Argoverse 2 motion prediction dataset with a reduction in speed distribution distance by 55.3% and off-road rate by 2.8%.
Robotic ultrasound imaging systems can improve acquisition consistency, but most existing systems only focus on normal positioning. A new framework integrates RGB-D perception, local-surface modeling, and task-space orientation control to enable omni-directional probe-orientation control. The system recovered a non-normal tilt angle of up to 44.39 deg relative to the surface normal and acquired desired heart chamber views in phantom and in-vivo experiments, with a mean angular tracking error of 1.06 ± 0.66 deg.
Empirically investigating human-robot interaction, researchers introduce an autonomy-supportive shared control strategy that preserves users' natural movement structure. In a user study, participants push an intelligent powered wheelchair under three conditions: no assistance, conventional variability-reducing assistance, and variability-preserving assistance. Preserving natural movement variability led to more favorable interaction experiences, with participants reporting significantly higher perceived agency and usefulness.
Researchers have introduced a "neuro-symbolic safety guard" to improve the safety and reliability of end-to-end autonomous driving agents. The guard, attached to the final command interface of an already-trained agent, checks explicit safety rules before executing commands. Evaluations on long-tail benchmarks show that the guard improves success rate by 15% and reduces safety-critical collisions by up to 53%, while preserving the original driving score.
Researchers Yanting Zhou, Jozsef Kövecses, and James Richard Forbes propose using the Koopman operator to represent nonlinear virtual environments (VEs) in kinesthetic haptic systems. Simulation and experimental results show that this method provides an effective representation of a Duffing-oscillator VE's nonlinear dynamics. The proposed method is more robust than traditional methods, as it uses a linear combination of lifted states to represent nonlinearity. A closed-loop stability analysis leveraging the Koopman representation confirms the overall haptic system's stability.
RoadWeaver, a coarse-to-fine framework, generates large-scale HD maps from scratch for autonomous driving simulation. It first synthesizes global road layout, expands it into connected network, and constructs lane-level geometry with topologically consistent connectivity. Experimental results show 99.8% reachability, 10.7% dead-end ratio, and 0.24m endpoint alignment error, reducing SOTA generation methods' errors by 94.4%. The generated maps can be deployed in driving simulators, providing scalable simulation environments for future closed-loop evaluation of autonomous driving systems.
A curriculum-guided shared policy learning framework is developed to enhance the navigation capabilities of autonomous maritime devices in smart ports. The framework uses centralized training and executes actions fully onboard, consistent with IoT edge intelligence paradigms. This approach improves navigation reliability, collision avoidance, and training stability compared to standard baseline methods.