Selecting a BLDC motor for robots must prioritize the mechanical system and load profile over the controller catalog specifications. Calculations involving torque, speed, and power, accounting for gear ratios and efficiency, are essential before controller selection. Control methods like Field-Oriented Control (FOC) offer advantages for smooth motion, while feedback selection depends on the required position accuracy and mechanical transmission. The choice of controller and feedback must integrate system-level concerns such as thermal limits, communication protocols, and comprehensive safety architecture. A structured workflow involving load definition, motion profiling, and system validation is crucial for optimal motor selection.
MIT engineers developed a soft robot powered by muscle cells that can swim by flapping fins in response to light. The robot's fins are covered with muscle cells genetically engineered to twitch when exposed to light. Flashing light on different fins allows the engineers to control the robot's direction and speed through a watery maze. The research optimized the underlying gel "skeleton" and muscle alignment to maximize the force generated by the muscle movements. This represents a significant step toward creating thin, efficient, and self-healing biohybrid robots.
Finding a dedicated solution for Over-The-Air (OTA) updates in the ROS 2 ecosystem proved challenging for the author. A personal project involving a Raspberry Pi attempted to implement an updater that pulled container images from GitHub and included a rollback mechanism. The author found that implementing a functional updater presented complexities beyond simply shipping Docker containers, noting issues with self-updating mechanisms. Approaches like A/B updates exist, but each introduces specific trade-offs. The author seeks input from production robot users regarding existing tools like Mender or balena for ROS 2 OTA.
The Robot Group, an authorized UK distributor for X-Humanoid, is soliciting expressions of interest from UK businesses and research organizations. They are seeking projects involving the company's TianGong and Tianyi humanoid robots for applications in manufacturing, logistics, hospitality, and infrastructure. The distributor will assist prospective customers in defining requirements and assessing suitable robot platforms. X-Humanoid's robots, including the TianGong Ultra, have gained international attention for their capabilities. UK organizations can engage with The Robot Group to discuss applications and explore these robotic solutions.
Origin raised $20 million in funding to scale its general-purpose construction robots, co-led by Accel and Zacua Ventures. The funds will be used to deploy its flagship robot, OG-1, and expand its fleet, supporting pilots across the US and Europe. OG-1 autonomously performs wall-finishing tasks, utilizing a Construction Action Model trained on real job site data. The company aims to automate interior trades by extending the robot's capabilities beyond wall finishing to other interior work. Origin plans to deploy 150 robots by the end of 2027 while working to increase construction speed and quality.
QCraft and HERE Technologies signed a memorandum of understanding to integrate HERE's map data with QCraft's navigation-assisted driving technology. This collaboration aims to accelerate the local validation and rollout of intelligent driving solutions in Europe and other international markets. Accurate, localized map data is crucial for navigation systems, which differ significantly across European countries. The partnership intends to enable safer, more reliable, and more localized driving experiences by combining QCraft's autonomy expertise with HERE's location intelligence. This move supports QCraft's strategy of providing "Global Technology, Local Service" for autonomous driving expansion.
Mercedes-Benz and Nvidia extended their strategic partnership to advance automated driving systems. This collaboration combines Mercedes-Benz's engineering expertise and MB.OS with Nvidia's AI and accelerated computing technologies. They are developing systems that integrate MB.OS with Nvidia Hyperion for in-vehicle computing and sensor architecture. Nvidia's Alpamayo Stack provides full-stack driving software, while Nvidia Halos enhance safety for advanced driver assistance. The partnership also extends to vehicle engineering simulation and the digital production ecosystem.
Agility Robotics plans to equip the Digit 5 humanoid with Nyobolt batteries, which allow for approximately 90 minutes of work after only nine minutes of charging. This performance yields a 10:1 ratio of working time to charging time, significantly improving efficiency compared to the Digit 4 model. This advancement aims to enable the robot to operate for over 20 hours daily, which will improve business economics. Nyobolt batteries are lightweight and designed to withstand numerous charging cycles. The company is also exploring methods like battery swapping for future humanoid models.
A Taku robot was deployed at Din Tai Fung in New York to fold napkins, according to Robostrategy CEO Andrew Kang. Kang suggested automating repetitive textile tasks in hospitality allows employees to focus on more skilled work. He noted that current robot deployments focus on specialized tasks like napkin folding, laundry folding, and data center operations. The development strategy involves building physical AI around constrained tasks to accumulate experience before achieving broader capabilities. Kang expects these specialized models to eventually generalize to a wider range of human activities.
China is investing in dexterous robotic hands as a crucial element for its humanoid manufacturing goals. Hand manipulation presents a major technical obstacle for creating functional humanoids, as coordinating grasping and multiple movements is technically complex. Government subsidies and a robust domestic supply chain provide advantages for this industry push. Companies like Linkerbot are developing this technology for various applications across different sectors. The development also addresses domestic labor needs by utilizing robots to fill workforce gaps.
DriveWorld3D proposes a framework for 3D world modeling and generative simulation to support real-world autonomous driving under computational constraints. It integrates temporal video learning, sensor measurement constraints, and offline generative simulation to create a unified world state. The system separates representations for physical occupancy, observability, and evidence provenance to manage uncertainty. Future prediction and risk-sensitive planning connect this world representation to autonomous driving decisions. The core hypothesis is that explicit evidence management can improve safety by reducing false free space predictions.
Mothership is a two-tier autonomous system for collecting refuse in pedestrian areas, addressing the difficulty for compaction trucks to reach bins on walkways. It involves an electric truck carrying two low-speed robots to fetch bins, return them, and recharge. Responsibilities are separated, with the truck handling driving and compaction, and the carts handling sidewalk transport and grasping. The system's performance depends on the interaction between cart count and stop-point radius, suggesting optimal configurations depend on density. Simulation results show that a two-cart configuration can achieve a throughput of 20-29 bins per hour.
GPT-6-Astra shows competence in robotic manipulation and navigation across various benchmarks. It demonstrates performance under in-context learning protocols involving demonstrations and extended execution steps. Close-loop coordination between the high-level reasoner and low-level controllers proves effective despite being costly. Real-world deployment reveals challenges related to collecting demonstrations and physical constraints. Overall, the results suggest GPT-6-Astra has achieved embodied competence but not yet practical embodied intelligence.
A framework called RAX addresses the limitation of existing robot learning methods by enabling agents to improve through online exploration and synthesizing executable programs. This is achieved by a general-purpose coding agent that generates robot control code without relying on demonstration data or value-function guidance. Through online exploration, agents discover goal-achieving decision sequences and compile them into modular programs encompassing the full perception-planning-control loop. A multi-embodiment study showed that generalization performance consistently improves as exploration scales across various systems. RAX autonomously learns complex manipulation tasks on real robots by composing verified skills, leading to emergent capabilities like error recovery and collaboration.
RTISSLAM-3D is a novel 6-DoF SLAM system designed for a wheeled robot equipped with an in-air 3D imaging sonar, an IMU, and wheel encoders. It utilizes a CFAR detector on the sonar images to generate sparse 3D detections, enabling 3D acoustic odometry and pose graph construction. This framework incorporates polar submap descriptors for place recognition and 3D registration for loop closure verification. Evaluation on a forest drive showed significant error reduction in loop closures when compared to 3D lidar SLAM references. The system successfully reduced horizontal error from 1.74m to 0.28m with a height error of 0.12m.
Long-horizon Vision-Language-Action (VLA) policies benefit from selective visual memory to manage computationally expensive observation history. RePlanMem-VLA uses policy plan revisions as a signal to guide the memory-writing process. It identifies task-relevant events by detecting large disagreements between temporally aligned future action predictions. This method successfully updates a bounded visual memory, preserving task-relevant observations in 100% of episodes. The findings show that plan revision can serve as an internal signal for controlling what a VLA model remembers.
High-precision global localization of mobile robots is demonstrated by utilizing natural ground texture viewed through microscope optics. The system relates optical magnification and image-space uncertainty to physical localization precision to address imaging challenges. The developed system achieved translation Root Mean Square Errors (RMSEs) of 28-73 $\mu$m across surface textures. Independent photometric registration matched feature-based localization estimates within approximately $9 \mu$m RMSE. This method shows that microscopic ground texture can enable mobile localization to achieve precision comparable to fixed positioning systems.
A lightweight embodiment classifier is trained to distinguish successful action chunks between source and target robot embodiments without requiring demonstrations. This classifier guides a frozen flow-based policy ($\pi_0$) by nudging its denoising steps with the gradient from the classifier. On the LIBERO benchmark, this guidance improved $\pi_0$'s average success on target bodies from 72% to 80%. This method is data-efficient, making it promising for sim-to-real transfer where rollouts are expensive. Guided policies also effectively address failures in specific scenarios, such as weak wrists, by allowing for the collection of additional target-body data.
FCTS is a failure-conditioned curriculum framework designed to guide legged parkour policies through training. It automatically generates the next training terrain based on localized traversal failures experienced by the robot. Monte Carlo Tree Search is used to explore potential curriculum trajectories within a fixed training budget. This method significantly improves the success rate on challenging terrains compared to existing baselines. Hardware experiments confirm that policies learned with FCTS successfully transfer to real-world scenarios.
Multi-robot authoring can speed up completion by distributing drawing work, but the speedup depends on task structure and the allocation-routing pipeline. The study compares single-token allocation with a shared-token pipeline, which allows multiple robots to work within a connected component. Shared-token execution showed larger mean speedups, though it did not reduce completion time for all drawings. Associations were found between structural descriptors and scaling behavior, emphasizing the need to consider within-component workload and graph structure. These findings highlight the importance of considering workload and graph structure alongside the component count for scaling in multi-robot systems.