Part 2/11:
Initially, training Optimus involved motion capture: humans wearing suits to demonstrate tasks for the robot to mimic. While effective for simple actions, this method encounters limitations when teaching complex movements like walking or dancing, because humans do not possess a robot’s balance or movement precision. To overcome this, Tesla has transitioned to using reinforcement learning in simulation environments, allowing Optimus to learn from virtual trials before moving into real-world application—a process known as zero-shot transfer. This approach significantly enhances scalability, reduces dependency on labor-intensive data collection, and accelerates the robot’s learning curve.