BallKick — kicking a ball

Mjlab-BallKick-Flat-MicroDuck

Kicking a 70 mm / 15 g ball forward. LB / RB kick left and right.

The policy cannot see the ball

This is the most interesting design decision in the task: the actor is ball-blind — the ball’s position is not in the observations.

So what it learns is not “aim at the ball and hit it” but “produce a stable, repeatable kicking motion”. Whether it connects depends on where you stand the robot.

The trade-off is sound: adding vision to the observations raises both the training difficulty and the deployment complexity by an order of magnitude, while a motion-stable kick is already enough to play with on hardware.

What that means for training

Because the ball is not observed, the reward can only be based on the motion itself — the kick’s amplitude, its speed, whether the robot is still standing afterwards — not on how far the ball flew.

A version that genuinely aims needs ToF or camera observations wired in, which is a research direction rather than a parameter change.