Wayve's self-driving taxis were added to the Uber app in London, allowing residents to hire these vehicles for the first time. Currently, only 15 vehicles are licensed, and all taxis still have a licensed human driver ready to take over control. The technology is based on an AI learning model, differing from rivals' mapping-based approaches. Wayve is working on scaling the technology and securing regulatory approval for fully autonomous services. This launch marks a significant step in bringing driverless taxis to London, despite ongoing discussions about full autonomy and social impact.
A new method called CW-Net translates the reasoning of autonomous vehicle AI into understandable concepts to explain its behavior. This technique uses concepts like "approaching stopped vehicle" to describe the vehicle's decisions, ensuring explanations accurately reflect the underlying actions. Researchers found that these explanations helped safety drivers better predict vehicle behavior in real-world tests. By providing clear feedback, CW-Net can improve human situational awareness and aid engineers in troubleshooting autonomous vehicle systems. Ultimately, this work aims to boost the safety and transparency of self-driving cars.
CJ Logistics deployed two ROBOTIS dual-arm robots on live Olive Young fulfillment lines in Yangji, South Korea. These semi-humanoids are used to stuff protective void-fill paper into parcels while human workers complete the packing process. Operational data from this deployment is used to continuously fine-tune CJ Logistics' Robot Foundation Model. The deployment emphasizes a shift toward pragmatic, wheeled semi-humanoids for complex logistics tasks in brownfield environments. This initiative integrates real-time operational data with AI to advance humanoid robotics in logistics.
Sam Altman confirmed OpenAI will definitely develop humanoid robots, arguing that matching human anatomy is essential because the physical world is built for people. While near-term focus remains on data center and industrial automation, the long-term ideal is ubiquitous domestic humanoids. OpenAI is building an internal robotics division and hiring mechanical talent to achieve this goal. Altman noted that non-anthropomorphic robots will be used for infrastructure, while humanoid form factors are favored for interacting with the physical environment. This commitment places OpenAI in direct competition with companies like Tesla and Boston Dynamics in the hardware race.
Minth Group and AGIBOT launched Europe's first mass-production facility for humanoid robots in Šabac, Serbia. This initiative involves an initial €20 million investment and aims to expand to a €200 million Robotics Industrial Park in Inđija. The goal is to achieve an annual production capacity of up to 20,000 humanoid robots and robotic dogs. The Serbian hub integrates manufacturing directly within Minth's automotive complex, allowing for real-time data collection for AI model training. This move establishes a European manufacturing base, creating a potential buffer against geopolitical trade restrictions.
X Square Robot introduced TwinDEX, a platform that pairs a wearable exoskeleton with a matched three-finger robotic end effector through physical isomorphism. This design enforces identical geometry, kinematics, and tactile sensing between the human operator rig and the robot, eliminating complex retargeting. TwinDEX allows policies to be trained on robot-free human demonstrations, achieving up to 5.3x data collection throughput compared to conventional teleoperation. The system successfully completed a complex chemistry experiment requiring only a few hundred demonstrations, with zero on-robot training data. This innovation addresses the embodiment gap by prioritizing data efficiency in embodied AI development.
A u-blox rover using 1 Hz GPS fixes reports a 5m horizontal 1-sigma position covariance, but the actual movement between fixes shows a median second difference of about 0.1 m. This discrepancy suggests the receiver internally smooths the data, making the fix-to-fix noise much smaller than expected for white noise. The author questions whether this smoothing is specific to u-blox or if other receivers exhibit similar behavior. They also inquire about practical methods for handling this smoothing in GPS applications.
The paper investigates the inverse kinematics of generic 3R robots using Conformal Geometric Algebra (CGA). It applies CGA to provide deeper insights into the kinematic properties of the robot. The method yields a univariate polynomial for $\theta_2$ by reframing the problem as the intersection of two circles within the CGA framework. This approach avoids the need to eliminate $\theta_1$ and $\theta_3$ algebraically. Ultimately, the work offers an alternative method for solving the inverse kinematics problem for these robots.
Kinematic analysis of a generic 3R robot is presented using a framework based on conformal geometric algebra. The paper develops a method to study the conditions under which a 3R robot possesses four inverse kinematic solutions (IKS). Necessary and sufficient conditions for a specific class of orthogonal robots having four IKS are derived using this approach. The work also allows for the study of the distribution of these IKS within the robot's workspace. Ultimately, the research provides a geometric analysis for understanding the kinematic properties of 3R robots.
The method extracts driving scenarios from egocentric datasets by generating a sequence of scene graphs from sensor data and HD maps. Linear Temporal Logic (LTL) is used to formally specify the driving scenarios of interest. An off-the-shelf model checker then extracts all instances of these specified scenarios from the scene graph sequences. This approach is effective for both simulated and real-world driving datasets. Evaluation on Argoverse 2 data demonstrates the method's success against a rule-based benchmark.
Autonomous overtaking requires decision-making that accounts for stochastic traffic evolution and temporal risk accumulation. A World Model-based Risk-aware Mixture-of-Experts (WM-RMoE) framework is proposed to address the limitations of standard safety approaches. This framework uses a learned latent dynamics model for multi-step rollouts and a hierarchical gating mechanism to coordinate safety modules. A Gaussian Mixture Model is integrated to manage multimodal maneuvering branches effectively. Experimental results show WM-RMoE improves safety compliance, stability, and generalization while generating diverse overtaking maneuvers.
Gaze-based manipulation for object selection does not guarantee accurate spatial placement, as pose errors can cancel each other out. The research introduces a markerless interaction framework and a cross-device dataset to study alignment from a task-oriented perspective. Graph-based Reference Selection is proposed to handle sparse robot references in this context. A new metric, Gaze--Surface Intersection Error (GSIE), is introduced to directly measure the spatial error of the gaze-specified target. Experiments reveal that alignment methods based solely on conventional pose metrics are not always optimal for gaze-based placement.
The research explores using nonlinear body oscillations to achieve natural quadruped gaits by leveraging mechanical resonance. The authors present a framework to make robot nonlinear dynamics predictable for design, allowing multi-gait locomotion to emerge from these resonances. They identify six nonlinear normal modes in a compliant quadruped robot named eBert using theoretical tools and simulations. Black-box optimization on these modes demonstrates how they naturally develop into distinct gaits with varying speeds. This work suggests a foundation for designing agile and efficient robots by harnessing embodied intelligence.
Language-driven grasping models fail when instructions include negation, such as specifying both a grasp location and an avoidance region. This work introduces an inference-time framework for negation-constrained dexterous grasping that does not require negation-specific training examples. It combines Sequential Monte Carlo with classifier-free guidance to steer sampling toward the grasp while pruning paths leading to forbidden regions. A frozen 3D part-grounding model is used to localize the forbidden area from the language instruction. Evaluation on the NegGrasp benchmark shows the method significantly reduces violation rates compared to baselines.
A system of two wearable pneumatic haptic devices facilitates continuous, closed-loop, bidirectional tactile interaction at relevant force and temporal scales. Each channel in a single device can function as a sensor, an actuator, or both, allowing for local closed-loop control of skin indentation. The distributed architecture supports rapid dynamic response and wireless pairing for real-time transmission and reproduction of tactile signals. Human studies demonstrated that this feedback improved teleoperated manipulation by reducing applied pressures and mental workload. The technology offers a practical foundation for bidirectional touch interaction by unifying sensing and haptic feedback through a single pneumatic modality.
AM-Bench is a modular simulation suite and benchmark designed for learning aerial manipulation policies. It addresses the unique challenges of aerial manipulation, including environmental disturbances and coupled dynamics. The suite includes various robot embodiments, twelve tasks, configurable disturbances, and various control and policy components. Unlike other benchmarks, it facilitates system-level evaluation of how embodiment, control, disturbances, and policy choices interact. The work validates this approach through simulations and real-world hardware testing.