Self-driving labs integrate AI with automated hardware to perform iterative experimentation. Unlike simple automation that follows scripts, these systems exhibit autonomy by learning from results in real time. They operate through a cycle: designing, making, testing, and then using the knowledge gained to design subsequent, more informed experiments. This allows the lab to intelligently navigate complex problem spaces by discarding unpromising areas of investigation. Essentially, autonomous self-driving labs enable machines to decide their next experimental step based on previous outcomes.
Hirebotics has introduced a new Cobraxt Painter, which is a no-code and explosion-proof collaborative robot (cobot) designed for painting applications. This system supports stationary painting processes. In operation, parts are positioned within a dedicated paint booth. Robots then autonomously handle all necessary spraying operations on these placed components. The announcement of this product was initially featured on The Robot Report.
Robust.AI is implementing the Aptiv PULSE sensor in its Generation 3 Carter mobile robot. This integration enables advanced sensor fusion capabilities. Specifically, the system combines radar and vision data using artificial intelligence. This technology allows the Carter robot to navigate safely around human presence. The announcement was initially published on The Robot Report.
RoboticsCareer.org has been expanded by the ARM Institute to include physical Artificial Intelligence (AI) roles. This updated platform now catalogs expanding job opportunities within the field of physical AI. Furthermore, it facilitates direct connections between employers seeking talent and qualified candidates. The expansion aims to better serve the growing demand for professionals in this domain. Coverage of this development was initially featured on The Robot Report.
Sony AI's Ace robot achieved a major milestone by defeating professional table tennis players in actual matches. The robot demonstrates advanced capabilities, tracking fast-spinning balls at up to 20 m/s and perceiving them within milliseconds. It operates with impressive real-time control, executing decisions on a 1 kHz loop. The system is notable for its successful integration and operation using ROS 2. This achievement showcases the potential of physical AI applied to high-speed competitive sports.
A tutorial is available for building a quadruped robot using ROS2 Jazzy. The author provides a GitHub repository containing the codebase for this project. This robot was constructed using an esp32 microcontroller and a Raspberry Pi. Users can access the necessary code base through the provided link. Feedback on the tutorial or the robot build is welcome from readers.
Industrial robots operating in demanding environments require extremely precise, high-speed coordination among multiple components like vision systems and microcontrollers. These sophisticated operations draw power from shared industrial power rails. The reliability of the entire system is threatened by sudden electrical surges or noise originating from other heavy machinery on the same grid. Such external disturbances can disrupt the delicate timing and control signals necessary for sub-millimeter precision movements. Therefore, addressing component-level power stability is critical for achieving more reliable industrial automation.
Building successful automation companies involves challenges beyond mere technical capability. Designing a functioning robot or automated process meets the assumed baseline requirement for these businesses. The significant gap lies in constructing the necessary surrounding business infrastructure. Many startups fail because they neglect this crucial organizational and operational scaffolding. Therefore, sustained success depends more on robust business foundations than just advanced technology alone.
Lab automation offers tangible benefits through robotic sample handling, automated liquid handlers, and comprehensive specimen tracking. These technologies significantly boost throughput while minimizing manual errors in diagnostic laboratories. As test volumes increase, automation helps alleviate pressure on stretched human workforces. The integration of these systems is no longer experimental but a proven, accelerating business reality. Therefore, improvements in lab workflow cannot rely solely on automation to compensate for deficient validation processes.
Autonomous rovers require enhanced obstacle detection beyond what traditional sensors or supervised methods can provide, especially beneath plant canopies. Video Memory Transformers for Anomaly Detection (VMTAD) introduces a fully unsupervised solution utilizing a transformer architecture with a memory module. This memory processes encoded representations from previous frames to account for dynamic scene changes caused by movement. Training relies solely on normal operational images, eliminating the need for labeled obstacle data. Evaluations on rapeseed datasets showed VMTAD achieved state-of-the-art performance (0.973 detection, 0.997 segmentation AUROC), with a lightweight variant maintaining crucial real-time inference speed of 14 ms.
Researchers developed a physically accurate simulation modeling blood capillaries, incorporating realistic fluid dynamics, red blood cells, and branching anatomy. Deep Reinforcement Learning agents were trained within this simulation to navigate using chemotaxis. The study systematically charted the physical boundaries of navigation based on robot size and swimming speed, identifying regimes where propulsion fails. Successfully trained agents independently discovered multiple efficient navigation strategies. These robots demonstrated the ability to autonomously block and unblock capillary flow without retraining, validating RL for in-vivo microrobotic applications.
Developing generalist robots capable of continuous task adaptation while leveraging prior knowledge is challenging. Previous methods using parameter-efficient fine-tuning fail to extract reusable skills or model skill interactions effectively. This paper introduces LiMoDE, which uses a novel two-stage learning scheme for lifelong robot manipulation. The first stage employs a dynamic Mixture-of-Experts (MoE) structure during pre-training to learn prior knowledge from heterogeneous experts based on motion information. Subsequently, the Lifelong MoE Adaptation Mechanism dynamically combines new and frozen experts for efficient knowledge transfer during task adaptation.
A novel, morphology-specific closed-loop control framework is presented for logarithmic-spiral continuum arms, which are suitable for mimicking biological appendages. This work models the system using a segmented tendon-driven approach in MuJoCo to account for compliance and contact dynamics. An analytical task-space Jacobian derived from spiral kinematics is enhanced with online compensation via Broyden secant updates and Kalman filtering to correct modeling errors. Validation through simulations confirms that this method significantly reduces tracking errors and suppresses attitude drift compared to baseline controllers. The resulting robust controller enables complex tasks such as obstacle-assisted grasping and multi-arm object handling for these underactuated manipulators.
Addresses the difficulty of learning long-horizon humanoid locomotion and manipulation by requiring both robust meta-skill execution and seamless chaining with autonomous recovery. Proposes OmniContact, a hierarchical framework built around "contact flow" (CF), which integrates key body trajectories and contact signals as a shared interface. A low-level policy, CF-Track, learns unified skills using this CF representation, while the high-level module, CF-Gen, synthesizes future contact-flow sequences for planning. The accompanying OmniContact dataset supports training on humanoid loco-manipulation. Evaluation shows superior performance over prior methods in both meta-skill execution and complex task chaining, with potential for VLM integration.
RoboTales is a low-cost robotic system that animates narratives using expressive sock puppetry. It autonomously synchronizes narration, gestures, and mouth movements, demonstrated on a Baxter robot. Pilot study results showed that puppet-based storytelling significantly outperformed gesture-only modes regarding HRIES ratings and story recall. This suggests that embodied puppetry boosts engagement and narrative comprehension in learners. Due to its modular design, RoboTales is adaptable for various manipulators, offering a screen-free educational alternative.
Teaches humanoid robots complex dynamic skills, like hitting a tennis backhand, using a coach-guided approach mimicking human instruction. The method, called TaskNPoint, requires the human expert to provide specific inputs: discrete skills, one demonstration per skill, the critical interaction window, and the overall goal. Training occurs in a physics-realistic simulation environment where unmodeled events are handled robustly. A key feature is randomized target sampling, enabling zero-shot generalization of a single demonstration to new locations. Testing on a Unitree G1 humanoid demonstrated successful learning of forehands, backhands, soccer kicks, and box manipulation using brief video inputs with minimal tuning.
Autonomous robot navigation requires accounting for surrounding human activities for safety and natural interaction in shared real-world spaces. Current research is limited by a scarcity of diverse, high-quality scene data in simulations. NavIsaacLab introduces a comprehensive framework to address this using physics-based and photo-realistic simulations of pedestrians and scenes. Built upon Isaac Lab, it utilizes GPU parallel simulation to provide accurate, real-time 3D visual feedback for robot training. To boost behavioral realism, the system incorporates a trajectory diffusion model and an adversarial motion learning controller for controllable pedestrian simulation.
A hierarchical control framework combining Model Predictive Control (MPC) and Reinforcement Learning (RL) is proposed for active roll control during quadruped racing. The system utilizes an MPC planner, based on a bicycle model of the Unitree Go2-W, to minimize lateral Load Transfer Ratio (LTR) in real time. A low-level RL policy controls the robot's 16 actuators, generating anti-roll torque via knee joints acting as active suspension during turns. Physical tests confirmed that this active roll control significantly reduces mean LTR by up to 44%. Furthermore, it improved the fastest lap time by 8.7% and increased peak lateral acceleration capability by 21.3%, maintaining high-speed stability.