Self-driving cars are becoming a reality, with Waymo already testing vehicles on London's streets despite some publicized mishaps. These automated vehicles (AVs) have historically required human backup drivers but will soon operate without them. Both Waymo and British competitor Wayve plan to launch driverless minicabs in London this year, pending necessary governmental approvals. For individuals with disabilities who cannot drive, these autonomous options present significant potential benefits. Operation in a complex environment like London presents a major challenge for the emerging technology.
Agility Robotics plans to become publicly traded by merging with Churchill Capital Corp XI. This corporate action is set to raise $620 million in capital. Funds raised will be specifically allocated to advancing the development of the Digit v5 humanoid robot. Furthermore, the capital infusion aims to help fulfill existing and growing customer orders for its technology.
Facing increased geopolitical pressure and the proposed GUARD Act, 1X is accelerating the release of its NEO humanoid robot. CEO Bernt Børnich announced this move to directly counter anticipated U.S. bans on Chinese robotics technology. The fast-tracking aims to address a significant hardware deficiency in the market. By speeding up developer access, 1X intends to maintain operational continuity despite potential restrictions.
Automation is transforming smart agriculture through increased use of tools like robotic systems, automated transporters, and advanced irrigation controls on farms and greenhouses. This equipment demands extended operational runtime, intelligent management, and robust performance in harsh outdoor settings. LiTime, a leading lithium battery manufacturer, addresses these industry needs with its specialized products. The company's offerings are designed to power the growing demand for reliable, long-lasting energy solutions in agricultural automation.
Version 2.1.0 of rclnodejs updates the package to be fully native ESM, facilitating its use as a web-native SDK for ROS 2. This allows developers to bring ROS 2 functionality directly to the browser and modern JavaScript toolchains. The updated structure offers typed APIs via `rclnodejs/web` for direct browser connection over WebSocket. Importantly, existing CommonJS projects remain backward compatible with zero code changes. This release represents a key step toward establishing rclnodejs as a first-class web SDK while maintaining access to the full ROS 2 runtime when necessary.
Bird's-eye-view (BEV) projection is a prevalent design pattern in applications like autonomous vehicles, robotics, and spatial AI. These models transform raw sensor data into a top-down perspective view for better scene understanding. Efficiently processing BEV information is critical for real-time physical AI performance. Developing faster pooling mechanisms specifically designed for NVIDIA GPUs can significantly accelerate this process. This optimization boosts the speed required for demanding, real-world AI implementations.
Mantis Robotics has introduced a new dual-arm, fenceless robot system. This innovative design offers increased flexibility for configuring automation processes. Its capability allows integration into complex, real-world industrial environments. Initial coverage of the launch was provided by The Robot Report. The robot represents an advancement in adaptable automation technology.
BitRobot, Hugging Face, and Unitree have collaborated to launch the Humanoids-in-the-Wild 500 (HIW-500) dataset. This substantial resource comprises a 500-hour teleoperation collection. The data was gathered within authentic residential environments in Southeast Asia. It was captured using Unitree G1 hardware, documenting real-world humanoid interactions. HIW-500 offers researchers vast material for training advanced humanoid AI models.
A system was developed to automate anomaly detection in ROS robotics by learning normal operational patterns without requiring manual threshold tuning or hardcoded rules. This system monitors a robot's topics and flags real-time deviations when something malfunctions. It has been successfully tested using public datasets covering faults on various hardware, including collaborative arms and quadrotors. The developer is now seeking beta testers to gain crucial experience monitoring the system on diverse, real-world robots. In exchange for free access and influence over development direction, the creator requests honest, detailed feedback from users in production settings.
Details an open-source project integrating the low-cost VL53L0X Time-of-Flight sensor into the Nav2 costmap as an obstacle source for small robots. The system pipeline connects the sensor to an ESP32, streams data via serial to ROS 2, converts it to a `LaserScan` message, and feeds it into the `ObstacleLayer`. Included components are custom firmware, a serial bridge node, and a converter that synthesizes a fan from the single beam. The implementation was tested on ROS 2 Humble and provides configuration advice for setup pitfalls.
A set of MIT-licensed React components named Altara provides real-time telemetry for robotics applications, including PFDs, moving maps, gauges, and event logs. These components are designed to be embeddable and themeable, allowing for a browser-based dashboard experience within existing applications. The author demonstrated building a small drone ground station utilizing these components by connecting them to MAVROS via rosbridge. This approach aims to replace the need for separate, native Ground Control Station software. The discussion invites others who have built ROS2 browser dashboards to share their topic subscription details and challenges with standalone tool limitations.
Introduces KRS Unleashed, an extension of the Kria Robotic Stack framework designed for easier hardware acceleration prototyping on FPGAs with ROS 2. The tool refactors the previous complex workflow into three distinct workspaces: 'os' for firmware and sysroot creation, 'vitis' to automate vendor-driven accelerator bitstream flows via Python CLI, and 'krs' for lightweight integration. This modularization significantly improves usability and build times, potentially speeding up iteration by up to 80x compared to older methods. Users can utilize these components in isolation or together for streamlined hardware-aware ROS 2 development.
Accurate extended pose estimation for multi-IMU articulated systems is challenging due to inter-body coupling and incorporating joint constraints. The authors introduce the relative L-extended pose, a Lie group representation allowing joint constraints to be expressed in an invariant form. This framework is utilized within an iterated Invariant Extended Kalman Filter (IterIEKF). IterIEKF incorporates these invariant constraints as noise-free pseudo-measurements while maintaining filtering consistency guarantees. Testing on a UR5e robot and human leg demonstrates that IterIEKF outperforms existing EKF and IterEKF baselines, achieving significantly lower RMSE and faster convergence.
Developing reliable autonomy in changing environments requires adaptive multi-sensor fusion for ego-motion estimation. The proposed ADM-Fusion is an end-to-end deep learning method designed to handle environmental changes and sensor failures. It uses an adaptive sensor mixture-of-experts framework with content-aware routing to dynamically weigh different sensor inputs. Furthermore, the system includes separate translation and rotation branches connected by a cross-task attention mechanism for specialized information sharing. Testing on CARLA-LOC and KITTI data confirms ADM-Fusion's robust performance despite degraded operating conditions.
Collaborative object goal navigation integrating robots with surveillance systems is crucial due to expanding deployment in indoor settings. Existing methods often fail when managing multi-view collaboration across large areas. Authors introduce SurveilNav, a framework designed to utilize multi-view surveillance information for systematic evaluation. The system combines active camera scheduling, joint 2D/3D mapping, and VLM-based value estimation. Experimental results on the HM3D dataset confirm that SurveilNav achieves state-of-the-art performance in efficiency and success rates by merging dynamic robot perception with global surveillance views.
Autonomous aerial vehicles require probabilistic methods to predict the unpredictable movements of non-cooperative obstacles like birds or debris. AeroCast introduces a framework combining a Transformer encoder with a Mixture Density Network (MDN) output head for this purpose. It predicts per-timestep Gaussian mixture distributions representing future 3D displacements, moving beyond single-point estimates. The model addresses input design and mode-degeneracy using consecutive displacement encoding and a calibration-oriented loss function. Tested on quadrotor data, AeroCast substantially reduces prediction errors while maintaining real-time inference speeds compatible with high-frequency onboard deployment.
Develops an RL guided Whole-Body Model Predictive Path Integral (MPPI) framework for humanoid control. The method utilizes a pre-trained Reinforcement Learning (RL) policy as a sampling prior to bias MPPI rollouts towards feasible behaviors, rather than using it as the final controller. Task objectives are integrated via modular MPPI cost terms, allowing online correction of the RL prior without retraining. Simulations on a 29-DoF humanoid show stable, high-rate control (averaging 280~Hz) in MuJoCo. This approach improves task precision by correcting systematic drift and tracking additional whole-body reference signals over pure RL methods.
Addresses challenges in safe control for autonomous driving involving multiple vehicles and varying behaviors. Existing methods struggle to model inter-vehicle influence causality or suffer from overly conservative worst-case analysis. This research extends Parametric-CBF by incorporating embedded causality inference to explicitly reason about interactions between surrounding vehicles. A resulting adaptive safety controller uses this causality-based function to allow the ego vehicle safe reactions based on learned neighbor expectations. Testing shows that utilizing motion flexibility significantly improves task efficiency in complex multi-vehicle scenarios.
Autonomous robots in partially observable environments need general memory retrieval mechanisms to recall diverse past information for long-term control. Incorporating transformer-based memory into imitation learning faces challenges, including learning spurious correlations and compounding memory errors leading to model drift. To address this, the proposed method, HALO, integrates an attention-based memory system for visuomotor policies. It suppresses false correlations by distilling vision-language model priors using video question-answering objectives on demonstration data. Furthermore, HALO mitigates accumulated error by employing sparse attention, allowing reliable retrieval from up to eight minutes of past experience.
SwarmFly is a new MATLAB-based simulation platform designed for designing and validating experiments involving Unmanned Aerial Vehicle (UAV) swarms. It addresses limitations in existing tools by offering a comprehensive, modular structure. The platform incorporates features like real-time operational maps, simulated IMU telemetry, and IP-based geolocation. SwarmFly supports four coordination modes and utilizes a plugin architecture for easy customization of behaviors and fault models. Testing confirms its capability to subject swarms to various disruptions while measuring metrics such as formation accuracy and fault recovery.
Low-latency inference in Vision-Language Models (VLMs) is crucial for real-time human-robot collaborative tasks. Weight pruning techniques can improve VLM efficiency for on-board processing while ensuring safe human-robot interaction. Safe interaction requires preserving "doubly-correct predictions"—outputs that are both accurate and evidentially grounded. The paper introduces a rationale-informed pruning strategy because existing methods often lose prediction accuracy despite retaining evidence localization. Benchmarks confirm this new method achieves superior prediction accuracy and doubly-correct performance on egocentric video datasets.