ASML, a Dutch company vital to the microchip industry, manufactures advanced lithography machines used to pattern minuscule circuitry onto silicon wafers. These complex machines are essential for creating increasingly powerful and dense chips needed for devices like phones and AI servers. The newest models utilize extreme-ultraviolet (EUV) light to achieve an unprecedented resolution of eight nanometers. Developing these machines required a decade-long, $10 billion R&D effort. The high cost of \$400 million per unit reflects the critical demand from chipmakers racing to improve chip capabilities and sustain Moore's Law.
A massive, silver, solar-powered platform is scheduled to traverse the southwestern US and cross the Pacific Ocean to Japan starting in August. This approximately 200-foot-long aircraft was constructed by Sceye, a company located in New Mexico. Upon reaching its destination near the coast of Japan, it will position itself about 18 kilometers above the ocean surface. The platform's deployment aims to improve internet connectivity from an aerial vantage point.
Drew Henry provided insights into the development of physical AI and robotics designed for real-world operational limitations. His discussion covered advancements in autonomous systems navigating complex environments. These strategies focus specifically on bridging the gap between advanced artificial intelligence and tangible, physical machines. The topic emphasizes building reliable AI for applications outside controlled settings. This deep dive was initially featured on The Robot Report platform.
ExRobotics has launched the ExR-2.5, an inspection robot designed for dangerous settings. This new model carries UL certification, ensuring safe operation within explosive atmospheres across North America. The announcement was initially published on The Robot Report.
Vention is partnering with major industry players, including FANUC and Universal Robots. These collaborations aim to simplify the process for manufacturers designing and deploying industrial and collaborative robots. By focusing on "software-defined automation," Vention seeks to streamline robotic integration. This effort makes advanced robotics more accessible to a wider range of manufacturing clients. The announcement was initially featured on The Robot Report.
Physical state recovery capability is identified as a crucial element for advancing physical AI. This ability will empower robots to effectively interact with their surrounding environment. The concept was highlighted in an opinion piece originally published on The Robot Report. It suggests that merely focusing on vision and language processing is insufficient for next-generation physical AI systems. Integrating accurate knowledge of the robot's physical state is necessary for practical advancements.
Agility plans an Initial Public Offering through a \$2.5 billion Special Purpose Acquisition Company (SPAC) merger. The transaction involves the participation of veteran dealmaker Michael Klein's SPAC vehicle. This move aims to list Agility on public markets. The company is positioned to capitalize on retail demand for humanoid robots. It is highlighted as the first Western pure-play in this sector.
Hyundai plans to take full ownership of Boston Dynamics through a $325 million buyout from SoftBank. This acquisition will establish Boston Dynamics as a wholly owned subsidiary of Hyundai. Following the purchase, the company intends to pursue a rapid listing on the Nasdaq exchange. The targeted timeframe for this Initial Public Offering (IPO) is by the year 2028.
Deploying large foundation models for robot control is challenging due to their opacity, which complicates formal verification despite their high expressive power. The paper introduces FEARL (Foundation-Enabled Assured Robot Learning), a framework addressing this by modularly decomposing the policy. FEARL separates a powerful Controller ($\text{C}$) from a small Safety module ($\text{S}$). $\text{S}$ operates on low-dimensional sensor observations and bounded context from $\text{C}$, enabling formal verification of critical safety constraints. Evaluation across simulated domains and physical robots demonstrates that this design preserves task capability while ensuring verifiable safety guarantees.
Engineering reliable autonomous systems is a crucial and expanding field in computer science due to increasing system prevalence. A recent workshop report summarizes discussions from the "Engineering Reliable Autonomous Systems" (ERAS) event held in June 2024. The workshop convened experts from formal methods, robotics, industry, and various application sectors. Research focused on three key areas: verification/validation techniques, engineering real-world systems, and safe software architectures. Ultimately, the report provides a roadmap detailing both existing but underutilized solutions and necessary future research directions for the field.
Reinforcement learning for bimanual piano playing often produces unnatural hand postures with high degrees of freedom. The proposed method, Adversarial Posture Regularization (APR), addresses this by using a small amount of casual human data instead of expensive expert demonstrations. APR incorporates an adversarial objective that forces the trained policy's posture distribution to match a learned human prior. The researchers also collected and released unstructured hand motion data from Meta Quest 3 users playing piano, retargeting it for the Shadow Hand. This approach significantly improves visual quality and adherence to human kinematics across multiple metrics compared to previous techniques.
Introduces Topological Online Learning for Displacement-based (TOLD) control, a real-time framework that adapts interaction topology weights directly to minimize formation distortion. TOLD proposes two methods: Online Gradient Flow (OGF) with unconstrained weights and Online Exponential Gradient Flow (OExpGF) with non-negative convex weights. Theoretical analysis proves OExpGF ensures asymptotic consensus for single-integrator agents on directed graphs, while OGF guarantees bounded distortion. Simulations show significant reductions in Root Mean Distortion Error when TOLD is combined with node-level controllers. Hardware testing using Crazyflie quadrotors confirmed substantial improvements in median formation distortion compared to fixed-weight methods.
Combines sim-to-real performance estimation with betting techniques and safe anytime-valid inference. The method utilizes scaled simulators to generate efficient and reliable certificates for estimating a mean value. This approach is particularly beneficial when conducting robot performance testing. The note provides a self-contained account of the construction, minimizing preliminary details by referencing original works. Synthetic examples demonstrating this proposed algorithm are available on GitHub.
Learning generalizable robot policies by scaling data across diverse robots faces challenges because actions are specific to the collecting robot's dynamics. To solve this, a new framework called SPACE uses Cartesian state delta as a universal action representation for all robots. SPACE operates in three domains: varying embodiments, different hardware units of one embodiment, and dynamic changes during operation. The framework comprises a Cartesian state delta policy predicting geometric end-effector displacement and an Action Adapter converting this into robot-specific commands. Experiments demonstrate that SPACE significantly outperforms direct control command prediction across varied robotics setups and maintains robustness to deployment dynamics shifts.
Imitation learning's data collection costs are addressed by trajectory-level data augmentation, which recombines expert demonstrations with varied initial states. Existing methods often introduce non-expert transitions through interpolation, potentially lowering data quality. Minimizing Interpolation (MinInter) introduces a selection method that chooses the source demonstration requiring the least interpolation for each sampled initial configuration. By explicitly minimizing interpolations during generation, MinInter creates higher-quality synthetic demonstrations compatible with current frameworks. Experiments on 12 tasks demonstrated MinInter consistently improves both data and policy success rates compared to methods like SkillGen.
DynaWM is a novel framework designed to improve bipedal-wheeled robot locomotion over continuous stairs, addressing limitations in current teacher-student control methods. It introduces a world model regularizer to enhance terrain encoding by enforcing forward-dynamics awareness and enabling visualization of terrain geometry. Furthermore, a momentum target encoder stabilizes knowledge transfer by providing consistent distillation targets against non-stationary updates. Analysis shows that the learned representations hierarchically capture terrain geometry with improved adaptability and smoothness. Experiments in both simulation and real hardware confirm DynaWM's superior performance on diverse continuous staircases.
Navigating complex environments presents challenges for conventional policies due to myopic decision-making and mode collapse. Existing world models often separate perception, generation, and control components inappropriately. NavWM proposes a unified model integrating latent world reasoning, multimodal action prediction, and controllable visual generation. It utilizes latent world tokens to encode geometric and semantic priors for structural understanding. This framework employs an anchor-based forecasting method generating diverse actions, enabling the model to function as a robust planner using visual foresight.