About the Project
Following on from the Booster T1's RL locomotion work, I trained a kicking policy so the humanoid could strike a ball on command instead of just walking. The goal was a controlled, repeatable kick that survives the jump from simulation to a real field, not a scripted animation.
Kicking is a harder control problem than walking. The robot has to shift its full weight onto one leg, stay balanced through the strike, and recover cleanly afterward, all while a real leg swings with real momentum and real contact with the ball.
Motion Retargeting
The kicking motion started as reference data that had to be retargeted onto the Booster T1's own joint limits and proportions before it could be used as a training reference in IsaacLab.
Sim-to-Real
The first working policy held up in simulation and transferred to the real Booster T1, real ball, real contact.
Vision as Input
This is the first policy I've built that takes anything like vision as an input. I don't simulate the camera itself in IsaacLab, the policy is trained on the ball's relative position as a privileged observation. On the real Booster T1, that same signal comes from its onboard depth camera, so the robot can aim and kick the ball as long as it can actually see it.