“Unitree G1” — still impressive.

This episode looks at the humanoid robot G1 by Unitree Robotics, which has been seen rallying tennis balls with human players.What’s especially interesting is how it was trained.The model started with just five minutes of low-quality human footage, learning basic movement patterns, and then improved through simulation using reinforcement learning.To keep the motion realistic, the system was constrained so it wouldn’t drift too far from human-like behavior.It also trained under varied physical conditions to better handle the gap between simulation and the real world.The result is a robot that can sustain rallies with strong accuracy and surprisingly natural movement.What stands out here is that progress no longer depends on large amounts of perfect data.With the right structure and engineering, even limited data can be enough to build something practical.It’s also a reminder that real progress often comes not just from theory, but from steady, hands-on engineering work.