Researchers led by Ganeshvar Soenessardien have used model-free reinforcement learning to teach humanoid robots such as the K1 to skateboard, without human demonstrations or a model of the board's steering, according to a paper on arXiv. The paper has been accepted at the 2026 IEEE-RAS Humanoids conference.
Training episodes end when the robot takes shortcuts, such as taking its planted foot off the board, standing on the board with both feet, twisting too far or leaving the board. How often each case occurs changes during training, which moves the robot from controlling its stance to riding the board continuously.



