AI Trainers Shift to Fully AI-Driven Tasks

"Our customers, ranging from large enterprises to AI research labs, are no longer just asking for AI features," said Anders Billesø Beck, VP of AI Robotics Products at Universal Robots. "They need a way to collect high-fidelity, synchronized robot and vision data to train AI models on the same robots they intend to deploy. Our AI Trainer is the industry's first direct lab-to-factory solution for AI model training." #IndustrialRobotics #PhysicalAI #AITraining #Automation #SmartManufacturing #ScaleAI #UniversalRobots https://lnkd.in/dKu_3daY

“Our customers, from large enterprises to AI research labs, are no longer just asking for AI capabilities,” said Anders Billesø Beck, Vice President of AI Robotics Products at Universal Robots. “They need a way to collect high-precision, synchronized robotic and visual data to train AI models on the same robot that will be deployed. Our AI Trainer is the industry’s first AI model training solution that goes directly from the lab to the factory.” #UniversalRobots #AIRobotics #GTC2026 #IndustrialAI #Automation #AITraining #Cobots #SmartManufacturing #ImitationLearning #FutureOfWork https://lnkd.in/d7FUTWNT

Robots learning skills through intention-based control suggest a shift from scripted actions to adaptive behavior. At Robotic Crew, we follow advances that bring autonomy closer to real decision-making. How important is intention for next-gen robotics? https://lnkd.in/dKJN9z_w

🤖 Universal Robots and Scale AI Bring AI Model Training Onto The Factory Floor A guided robot setup enables real-time data collection for training Vision-Language-Action models using the same machines deployed in industry. Read more: https://lnkd.in/g8hnc_YT #physicalAI #VLA #AI #robotics

🤖 Teaching robots how to recover from falls using AI In real-world environments, falls are inevitable for legged robots navigating uneven and unpredictable terrain. Enabling reliable recovery without interrupting operation remains a major challenge due to complex contact dynamics and external disturbances. A recent study published in Biomimetics introduces a deep reinforcement learning-based framework for dynamic fall recovery in quadruped robots. 🔬 Key innovations: ▪️ A learning-based state estimator capable of handling complex and uncertain contact conditions ▪️ A proprioceptive history policy that leverages past sensory data to guide recovery actions ▪️ A unified framework applicable to diverse fall scenarios, both indoors and outdoors 📊 Key outcome: Extensive experiments demonstrate that the learned policies can be successfully deployed on real robot hardware, enabling effective recovery from falls on both flat surfaces and grass terrain. 💡 This work represents a significant step toward more autonomous, resilient, and field-ready legged robots. 📖 https://lnkd.in/dg2D_AdR #Robotics #ArtificialIntelligence #ReinforcementLearning #QuadrupedRobots #Biomimetics #AutonomousSystems #Innovation

Most robots today work with one arm. Humans don't. Think about folding clothes, cooking a meal, opening a jar, assembling furniture, or handing something to another person. Almost every meaningful task in daily life requires two hands working together. That's why bimanual manipulation is one of the most important frontiers in robotics right now. It's the bridge between "lab demo" and "robot that actually helps you at home." We just crossed that bridge at Vizuara. Using two pairs of SO-101 robot arms (4 arms total — 2 leaders for teaching, 2 followers for execution), 3 cameras, and SmolVLA — a compact Vision-Language-Action model — we trained a single policy that coordinates both arms to perform a handover task: (1) The right arm passes a bowl containing a box to the left arm (2) The left arm picks up the box and places it in the correct bowl (3) All from a natural language instruction One model. 12 degrees of freedom. Zero hardcoded logic. The model understands what to do from language ("pass the red bowl") and figures out the bimanual coordination entirely from 10 human demonstrations. That's it. 1 hour of training on a single A100, and the robot generalizes. Here's what makes this real: - Runs on a Mac Mini with Apple Silicon — no cloud GPU needed for inference - Uses open-source everything: LeRobot framework, SmolVLA model, SO-101 arms - The entire pipeline (recording, training, deployment) works end-to-end Why this matters for Vizuara: We're not just teaching robotics theory. Our students in the Modern Robot Learning Bootcamp are building and deploying these systems on real hardware. From imitation learning to VLA models to now bimanual coordination — this is the progression from classroom to cutting edge. The full pipeline, code, scripts, and demo videos are open-source: https://lnkd.in/dT3KKeDk Trained model on HuggingFace: https://lnkd.in/d-XR8HzD We are getting closer to robots that can do laundry, cook, clean, and assist the elderly. Not in 10 years. Now. One task at a time. Link to our bootcamp: https://lnkd.in/dhH4wxzj

Universal Robots launches AI Trainer for factories Universal Robots is moving AI training onto factory-grade robotics hardware. The UR AI Trainer combines leader-follower cobots, force feedback, and Scale AI software to capture production-ready datasets for model training and deployment....

Posts on humanoid robots have been inundating my social media feed, so I decided to explore the subject a little further. This post talks about three robots currently in use; in later posts, I will be discussing different, human-related aspects of training them. Here are three humanoid robots that are defining how robots will look in the near future. Figure 02 is the one with the strongest industrial track record. Partnered with OpenAI and deployed at BMW's Spartanburg plant, it moved over 90,000 components and logged roughly 1,250 operating hours across its pilot run. It understands voice commands and can execute tasks in real time. With more than 40 degrees of freedom, it is built for precision in high-volume manufacturing. The catch: you cannot buy one. It is currently only available to selected research and industry partners. Tesla Optimus takes the mass-market approach. At 1.73 metres tall, with a payload capacity of over 20 kilograms, it is designed for realistic industrial work. Tesla's strategic advantage is its AI infrastructure, with neural networks from its self-driving vehicles being transferred directly to the robot. Over 1,000 units are already deployed in Tesla's own factories. The target price sits between $20,000 and $30,000, though independent validation of its full capabilities is still pending. Unitree G1 is the accessible disruptor. Starting at $16,000 and available for global order today, it is the only humanoid in this comparison that researchers, universities, and smaller companies can actually get their hands on. It is compact, remarkably agile, and runs on an open software architecture that has attracted an active developer community. Its limitation is payload, at 2 to 3 kilograms per arm, making it a platform for experimentation rather than heavy industrial automation. (P.S.: Unitree has announced it will be launching its cheapest humanoid robot priced at around $4000 via AliExpress this week). However, significant challenges remain. A recent paper published in Nature's Scientific Reports outlines what still stands between these machines and real-world deployment: battery life that rarely exceeds one to two hours of continuous operation, payload-to-mass ratios that make heavy material handling impractical, sensors that degrade in dust, glare, and bad weather and regulatory frameworks that have no clear answers for who is liable when a robot causes an accident on a construction site. Many of these systems still rely on external motion-capture setups or pre-planned trajectories, which is a far cry from the adaptability that messy, unpredictable environments demand. #Robotics #AI #HumanoidRobots #Automation #Tesla #Figure02 #Unitree

Over the past few decades, robotics researchers have developed a wide range of increasingly advanced robots that can autonomously complete various real-world tasks. To be successfully deployed in real-world settings, such as in public spaces, homes and office environments, these robots should be able to make sense of instructions provided by human users and adapt their actions accordingly. #Robotics #ArtificialIntelligence #NaturalLanguageProcessing #MachineLearning #Automation

A new framework has been introduced that integrates large language models (LLMs) with the Robot Operating System (ROS) to enhance natural-language control of autonomous robots. This approach enables robots to translate user instructions into executable actions, supporting both inline code and behavior tree execution modes. Initial experiments demonstrate robust performance across diverse tasks and robot types, utilizing open-source pretrained LLMs. The framework is available as open-source code, offering opportunities for further development and application in complex, real-world environments. This advancement represents a significant step toward more adaptable and responsive robotic systems.