Debby

A new empirical study from @axisrobotics examines how simulation data can help robots learn with very little real-world training data and why the quality and selection of that data matters. The researchers tested three manipulation tasks on a physical Franka Research 3 robot: StackCube, Peg, and CupCake. Each task started with just 10 successful real-world demonstrations, alongside a pool of 1,200 simulated demonstrations. The results: • Real data only: 16.7% average success • Full simulation pool: 25.8% • Random 600 simulations: limited and inconsistent improvement • Selected 600 simulations: 85.8% average success That means carefully selected simulation data improved real-robot performance by 69.2 percentage points compared with using the same 10 real demonstrations alone. So what made the difference? Axis used TUCO, a data-curation method that uses influence functions to identify demonstrations that are useful and complementary to the target task. The selected simulation data wasn't simply more data. It was more relevant data. The experiments covered grasping, transfer, alignment, stacking and contact-rich insertion, with the robot learning from three camera views alongside its joint and gripper information. The study also tested performance under unseen lighting. Selected simulation data still performed better than real-only training, but the difference wasn't statistically significant in that test. So the results shouldn't be taken as proof of general robustness. The bigger takeaway is simple: Scaling robot-learning data isn't necessarily enough. The composition, relevance and selection of that data can determine how effectively simulation translates into real-world performance. If robots can learn more from carefully selected simulation data, we could potentially reduce the amount of expensive real-world demonstration data needed to train capable physical robots. That makes sim-to-real learning not just a question of how much data we have, but how intelligently we use it. An important step toward making robot learning more data-efficient and practical to scale. s.kaito.ai/tz1O3q7

There’s more to @ActionModelAI than just a browser extension. The ecosystem is being built around three products: Action Model, ActionFi and Actionist. Action Model is the foundation. Its browser extension allows people to contribute real-world browser interactions that help train a Large Action Model to understand not just what humans say, but how they actually use computers to get things done. Your actions become training data. Your contribution gets rewarded. And the community helps build the intelligence. ActionFi puts that intelligence into a real incentive layer. Instead of rewarding likes, follows or surface-level engagement, ActionFi rewards people for completing meaningful actions on real products. Projects define the actions that matter. Users complete them. The browser extension verifies the journey. Real users. Real product usage. Verified actions. Rewards. Actionist is where that intelligence becomes useful. It takes computer-based work and turns it into AI-powered automation. From research and lead generation to updating CRMs, managing workflows and other repetitive tasks, Actionist is designed to let AI actually perform work rather than simply tell you how to do it. So the three pieces fit together: Action Model learns how humans act. ActionFi rewards and verifies those actions. Actionist turns that intelligence into automation. That’s what makes the ecosystem interesting to me. It isn’t just about building another AI product. It’s about creating an AI economy where people can contribute to the intelligence, use the technology and participate in the value being created. Train. Earn. Automate. Own. Join here: join.actionmodel.com/HOTTIEB…