Waymo is integrating Gemini into its custom Ojai vehicles to provide a more personalized ride experience. Gemini is an in-car AI assistant that can be controlled using voice commands, allowing riders to control the cabin, learn about their surroundings, and access information on nearby attractions. Gemini operates independently of the Waymo Driver and is only active when engaged by the rider.
NVIDIA Cosmos 3 Edge is a 4B omnimodel with a 2B NVIDIA Nemotron-based reasoner that has been pretrained on the same physical-world data as NVIDIA Cosmos 3 Nano and NVIDIA Cosmos 3 Super. It starts with the same grounding in how objects move and interact and is small enough to run on-device on NVIDIA Jetson Thor. The model can be used to predict robot actions and has been evaluated in closed-loop simulation, reaching 22.9% success across 120 language-conditioned manipulation tasks.
The second World Humanoid Robot Games (WHRG) kicks off on August 22 in Beijing, with 2,056 robots from 666 teams across 16 countries participating. The event has shifted focus toward commercial utility, introducing 21 scenario-based events that test robots in simulated factory, hotel, and retail environments. Autonomy rules have been tightened, with task time limits slashed and remote-controlled operations facing severe scoring penalties. The event aims to test whether humanoid robots can perform reliable, autonomous work to generate economic value.
MIT engineers have designed a new, more intuitive controller for operating construction diggers, which uses a miniature excavator's arm and bucket to mimic the machine's movements, eliminating the need for mental mapping and potentially shortening the training process for operators. The new interface, named the "World-Space Interface," pairs the mechanical arm with a virtual simulation of an excavator, allowing operators to physically operate the machine in a virtual environment. The team reports that novices using the WSI were just as good as experts from the start, compared to the traditional joystick-based simulator, which required novices to learn complex mental maps.
A researcher has developed a runtime called xZane that sits between an LLM and physical actuation, validating commands against device capabilities before execution, and running independently of the model. The approach has been tested on various devices, including a mobile robot, LG webOS TV, and robot vacuum, with similar results across different models and environments. The xZane runtime differs from related work, such as ROSClaw and URML, by integrating a device's existing safety function and avoiding the safety-critical path of the model. The researcher is seeking feedback from others to determine if the approach fits their use cases.
VirtuLayer is a web-based GUI developed to simplify the configuration and usage of the Nav2 Virtual Layer Plugin. It provides a visual interface for creating virtual geometries on a map and generating the required configuration. The plugin can be used to add virtual geometries to the Nav2 costmap, allowing for a simple visual workflow for creating and experimenting with virtual layers.
ROS 2 Inspector is an open-source project aimed at automating the architectural pass for ROS 2 workspaces. It uses static analysis to provide a compact summary of the discovered ROS 2 packages and their health signals. The tool generates interactive visualizations and can export data in table, JSON, or YAML formats. ROS 2 Inspector is not intended to replace runtime tools, as dynamic names and behavior can only be fully understood with runtime evidence.
X Square Robot is showcasing its full portfolio of embodied AI technologies at the World Robot Conference 2026 in Beijing, presenting applications in logistics automation, dexterous manipulation, household robotics, and AI training-data collection. The company's WALL-B embodied AI foundation model, robotic hardware, manipulation technologies, and data infrastructure are on display at booth C107. X Square Robot's demonstrations include a logistics system that processed 1,816 parcels with an accuracy of over 98 percent and a flower-arranging demo that tests the robot's ability to understand natural-language instructions and perform complex manipulation tasks.
Automated T-shirt printing machines are revolutionizing the textile industry by supporting businesses in printing more shirts in less time, making the work easier and minimizing errors. These machines can print various designs, including company logos, pictures, names, numbers, and colorful designs, producing clear, vibrant, and neat prints. They offer several benefits, including faster printing, better print quality, less manual work, fewer mistakes, and lower production costs.
Global humanoid robot shipments exceeded 22,000 units in the first half of 2026, with AGIBOT leading the market with 9,700 shipments (43.1% market share), followed by Unitree with 7,000 units (31.1%). The top five vendors controlled over 86% of the market, with entertainment, performance, and data-collection research accounting for over 60% of deliveries.
Regular price ticket sales for ROSCon Global end on Monday, August 24, 2026, at 12:00 AM UTC. After-hours events, including an OSRF hackathon and a bus tour, will be announced later next week. Registration for ROSCon Global is required to participate in the after-hours events, with limited space available.
Authors developed a mixed-integer programming model to maximize value retention in wildfire-threatened regions through scheduling and routing of unmanned aerial vehicle (UAV) fleets. The model considers degradation-triggered job arrivals, where new jobs appear when the demand of the existing job surpasses a threshold. The overall objective is to maximize the total remaining reward, aligning with the proverb "a stitch in time saves nine." The model is scalable, and its performance is evaluated through computational experiments and a case study.
Researchers Md Mahmudul Islam and Shaurya Agarwal propose a trust-field framework that aggregates microscopic vehicle-level trust into a continuous representation over space and time for vehicular networks. The framework is formally defined on road segments and enables analysis of trust evolution across road segments through simulation-based experiments. A field-informed deep learning method that treats trust as a latent quantity carried by vehicles and enforces measurement consistency through aggregation yields improved reconstruction error.
A robot is prompted to create an algorithmic solution to the "Push-T" task without human demonstrations, using an LLM coding agent. The agent learns push mechanics using a 2D gym simulation and achieves a 100% success rate with 46% fewer steps than a baseline. The agent also solves extensions of the task using a self-generated curriculum and visual feedback.
GigaBrain-WBC-0.5 is a Behavior World Model for humanoid whole-body control, enabling robust interaction with environment and staying robust to implausible commands, falls, and disturbances. It achieves the highest success rate across all four regimes among three large-scale tracker baselines. The model achieves 81.3% success rate on terrain interaction, 83.1% under implausible commands, and 99.3% fall recovery. Hardware trials show robust interaction under missing supports and disturbances.
GAPL (Grounded Action-effect Policy Learning) is a unified framework that integrates Large Language Models (LLMs) with simulation-based effect grounding and policy optimization. The framework consists of three modules: an LLM-based Effect Evaluator, a Simulation-based Effect Grounder, and an Effect-Aware Decision Maker. Experiments on four Highway-env scenarios demonstrate that GAPL outperforms baselines, achieving average reductions of 0.76, 0.86, 2.00 in collision rate, average displacement error, and final displacement error, and an average reward gain of 1.44.
VERAGMIL, a framework combining a high-fidelity simulator and a virtual reality interface, provides a realistic environment for training robot-assisted feeding systems to handle granular materials. The framework is evaluated with three imitation learning models, achieving the best overall performance, particularly in reducing spillage, and approaching human performance. The code for VERAGMIL is publicly available.
A task-agnostic control strategy for dynamic assistance with pneumatically actuated soft exosuits was developed, relying only on kinematic sensing. The strategy uses a Hammerstein dynamic model to capture static and dynamic behavior of the human-robot system, and approximates an inverse to integrate into the control loop. The controller reduced interaction torque by up to 73% and activation of key muscles by up to 47% when evaluated on a test rig. The control strategy provides task-agnostic, dynamic assistance without the need for physiological or force sensors to interpret user intention.
Robot foundation models, including vision-language-action policies, often focus on scaling up data, models, and benchmarks, but practical deployment on a target robot requires additional work. This work is categorized into the "embodiment gap," which includes shared structures and adaptation needed for execution on a new robot. Recent work addresses sharing semantics and perception, robot data and interfaces, and learning correspondence across embodiments. A reporting framework for adaptation work highlights the need to investigate what should be checked when comparing cross-embodiment learning.