A European fertility organization is calling for international limits on the number of children a single sperm donor can contribute to. This concern stems from experiences where individuals discovering their parentage through anonymous donors have found numerous half-siblings. These findings prompt advocates to address potential issues related to feeling "mass-produced." The proposed limitations aim to manage the implications of multiple biological connections traceable to one source.
China is extensively adopting artificial intelligence across numerous sectors. Examples of this embrace include medical consultations with AI avatars and food delivery via drones on structures like the Great Wall of China. Intelligent robots are also being utilized within manufacturing facilities nationwide. Furthermore, the state government has eagerly incorporated AI technology. This integration particularly highlights new capabilities for advanced surveillance measures.
STMicroelectronics has acquired a stake in Oversonic Robotics, an Italian cognitive robotics firm. Oversonic Robotics specializes in developing humanoid robots, notably RoBee, which is the first certified model for complex environments. New investment from STMicroelectronics, Fondazione ENEA Tech Biomedical, and SpotInvest aims to boost the company's growth. The funding will support Oversonic's industrial, technological, and international expansion efforts.
Shipping errors frequently occur during routine handoffs, such as moving data from an order screen to packaging and then into a carrier system. Time pressure exacerbates these manual transfer points. Modern automation solutions minimize the necessity for staff to manually retype or interpret information at these stages. By streamlining data flow, these technologies drastically reduce human error potential. This improves overall accuracy in the shipping process.
Wristbands allow wearers to operate a robotic hand using natural hand and finger movements. This control mechanism enables tasks such as playing the piano with the robot. Furthermore, users can aim the robotic arm to simulate shooting a basketball. The technology also facilitates object manipulation within virtual environments. Overall, personal movement directly governs the robotic hand's actions.
Robot.com has launched R-noid, a new humanoid robot designed for demanding labor roles. This machine is specifically intended for repetitive, multi-shift, and hard-to-staff jobs. The company operates under a Robot-as-a-Service (RaaS) model for deployment. Customers can see the robots function from an initial site visit to autonomous on-site operation. This launch marks Robot.com's entry into providing humanoid labor solutions in real-world settings.
AgenticFocus is a Mixed Reality synthesis pipeline designed to transform standard first-person human videos into robot-trainable demonstrations for humanoid learning. It addresses limitations in existing pipelines by restoring occluded object geometry and reconstructing detailed hand motion from challenging video inputs. The method achieves this through camera-relative alignment and layered compositing techniques, enabling accurate retargeting to a humanoid body model. This process generates a dataset pairing focused visual observations with synchronized robot actions and states. Experimental results show AgenticFocus yields lower trajectory error and smoother wrist motion compared to existing cross-embodiment baselines.
Generative robot policies struggle with novel scenarios outside their initial training distribution, making safe real-world adaptation difficult. FlowDAgger offers an efficient method to adapt these frozen policies using minimal human demonstrations directly within the latent space. It achieves this by inverting expert actions into corresponding noise inputs via reverse-time integration and local refinement. This resulting inverted noise trains a lightweight policy that guides the base model during deployment for rapid skill acquisition. Evaluations show FlowDAgger surpasses supervised fine-tuning and latent-space RL baselines while maintaining existing capabilities on separate tasks, providing practical adaptation for foundation models.
SplatCtrl is a novel framework designed for collision-free robotic arm control in unstructured and changing environments. It utilizes 3D Gaussian Splatting (3D-GS) combined with a hybrid voxel-based filtering and dynamic Gaussian relocation strategy for efficient real-time scene reconstruction from RGB-D data. The method develops continuous signed distance functions from isotropic Gaussians, offering stable collision probability estimates. These metrics integrate into control barrier functions to ensure safe, reactive motion planning. Testing in simulation, physical robots, and shared workspaces confirms its success in integrating perception and action within dynamic settings.
Dec-MARVEL is a decentralized framework designed for multi-UAV exploration under constraints such as unreliable communication and finite travel budgets. It enables coordination without explicit communication by utilizing incidental observations, treating any teammate visible in the field of view as a signal. The system employs a graph-attention actor that fuses local frontier geometry, observed teammate motion, and budget status to select safe waypoints. Training involves phase-conditioned critics and a mixture-based budget curriculum. Evaluations across various team sizes and budgets show Dec-MARVEL achieving superior exploration rates compared to baselines, with successful physical deployment demonstrated in real-world tests.
Accurate dynamics modeling of Brushless DC (BLDC) motors is crucial for high-performance robotic joint control. A Physics-Informed Neural Network (PINN) utilizing a deep residual backbone was developed to create a continuous-time surrogate model of the full six-state BLDC motor dynamics. Given inputs like voltages and excitation parameters, the network predicts all state variables—rotor angle, angular velocity, currents, and temperature—while enforcing physical laws via a combined loss function. The training process employs curriculum scheduling for stability, finishing quickly on standard CPU hardware. Post-training inference achieves very low latencies (0.1–22 $\mu$s), demonstrating up to 118x speedup over conventional ODE solvers for real-time control.
Autonomous driving systems struggle when key actors are temporarily hidden due to occlusions, causing standard models to lose track of these agents. BeyondSight is a novel framework addressing this by decoupling actor existence from immediate observability using persistent hypotheses. It temporally propagates and updates these hypotheses with available evidence, allowing joint perception, prediction, and planning for unobservable actors. To support rigorous evaluation, the authors introduced nuScenes-Permanence, an extension providing supervision for hidden agents. Experiments confirm that BeyondSight improves reasoning under occlusion by boosting detection performance and reducing planning errors compared to existing methods.
Human-Object Interaction requires tactile feedback to achieve realistic dexterity, but current methods often neglect physical contact during robot imitation. Introducing TactiDex, the authors provide a new, real-world benchmark dataset linking whole-hand tactile signals with kinematic and object states. This framework enables moving dexterous tasks beyond simple movement mimicry toward accurate contact-level behavior. A proposed method, TactiSkill, utilizes a novel tri-component tactile reward to guide robotic execution using human demonstrations. Experiments confirm that TactiSkill outperforms others in both manipulation success and physical realism for single and bimanual tasks.
Generated videos can guide robot manipulation but often lack physically actionable metrics like precise geometry or kinematic feasibility. The proposed GenVid2Robot framework converts these video motions into executable trajectories by enforcing rigid-geometric consistency. It samples semantic anchors from the initial frame and verifies if subsequent 2D motion is explained by the first frame's RGB-D geometry under an $SE(3)$ model. This process treats videos as uncertain hypotheses, transferring only geometrically consistent relative motion to grasp planning. By integrating this with mask-constrained grasping and a depth-compensation module, GenVid2Robot generates reliable execution trajectories for real-world robotics.