Moonlake unveils simulation method that boosts robot learning success by up to 58 points

Moonlake outlines a "Calibrate-Evaluate-Improve" loop for building robot policies with simulation. Simulations are calibrated using little real-world data via real-to-sim, physics modeling, and inverse physics calibration, providing strong sim-real correlation for policy evaluation. Evaluation results guide domain randomization, followed by rejection fine-tuning and RL to improve policies, which the company says transfer to real robots. The approach also enables online learning, with a robot using physics simulation as its world model and taking actions that reduce world-model uncertainty over time. Moonlake invites partners to accelerate policy evaluation and deployment.

✨ AI summary of the linked source — not editorial content.