Episode Details

Vision-language-action (VLA) model enabling zero-shot sim-to-real transfer for mobile manipulation tasks, trained entirely in simulation without real robot data, achieving 79.2% success on real-world DROID benchmarks outperforming baselines by 2x.