← Paper Digest

Reference-Free Humanoid Skateboarding via Termination-Shaped Reinforcement Learning

Ganeshvar Soenessardien, Michael Przystupa, Kevin Sebastian Luck

November 2026

TL;DR A Booster K1 humanoid learns in MuJoCo to push itself along on a four-wheel roller board with PPO and no motion reference, guided only by task rewards and early terminations that rule out exploits. Across three seeds it holds the one-foot-on-board push stance 97.6% of the time and moves the board at 0.27 m/s.

What they do

Earlier humanoid skateboarding (e.g. HUSKY) relied on human pushing demonstrations, an analytic steering model and sim-to-real engineering. This paper asks how far plain PPO gets without any of that: no motion capture, no target joint postures, no posture-tracking reward. A simulated Booster K1 (MuJoCo via mjlab) must propel a passive four-wheel roller board forward while staying upright.

How it works

  • Push stance. Each episode starts with one foot on the deck and the other on the ground beside it (the “XOR stance”). The actor is warm-started from a generic walking policy; commands ask for 0.2-0.6 m/s forward.
  • Rewards. Forward board motion (the largest weight), forward trunk motion counted only in the XOR stance, uprightness, trunk height and keeping the off-board foot low; penalties for standing idle on the deck or moving backward.
  • Exploit-blocking terminations. Episodes end if the board foot slides more than 5 cm, both feet are on the deck, body-board yaw exceeds 0.4 rad, the trunk drops below 0.40 m or the robot leaves the board. Board-foot displacement dominates early and board out-of-bounds later, an implicit curriculum from stance control to locomotion.
  • Compute. 6,144 parallel environments on one RTX 5070 Ti; a 20,000-iteration run takes about 28.7 h.

Results

  • Over three seeds, final board speed is 0.270 ± 0.087 m/s and trunk speed 0.308 ± 0.095 m/s, so the robot travels with the board instead of kicking it away.
  • The XOR-stance rate passes 0.97 by about iteration 500 and ends at 0.976 ± 0.009; episodes last 498.8 ± 76.5 of 1,000 steps.
  • Seeds differ: board speed ranges from 0.185 to 0.359 m/s, and the fastest seed has the most out-of-bounds terminations.
  • The stroke is conservative: the pushing foot stays near the ground (near-ground rate about 0.92), a small lift-and-replant rather than a full swing.

Limitations

Simulation only, and the actor observes board state that real hardware would need an estimator for. The board has no trucks, so there is no lean-to-steer turning, only straight pushing on flat ground. Three seeds only, and neither the terminations nor the near-ground reward were ablated, so their individual effects are not established.

Figures

From Figure 1: A K1 Booster humanoid learning to skateboard on a mobile platform with model-free reinforcement learni...
From Figure 1: A K1 Booster humanoid learning to skateboard on a mobile platform with model-free reinforcement learning, front and back view.
From Figure 6: Side view of the propulsive stepping sequence, from load/contact through early push, toe-off, foot lif...
From Figure 6: Side view of the propulsive stepping sequence, from load/contact through early push, toe-off, foot lift and repositioning to re-contact.
From Figure 7: Overhead (left) and oblique (right) snapshots of the robot and board moving through the simulation; th...
From Figure 7: Overhead (left) and oblique (right) snapshots of the robot and board moving through the simulation; the orange line traces the board centre from the green start marker.
Abstract

Humanoid locomotion research has advanced rapidly on walking and running, but \emph {object-mediated} locomotion, in which a robot must coordinate its body with a passive external object, remains difficult. One challenging task is skateboarding, which requires simultaneous interaction with the ground and a passive but movable platform to achieve forward motion. Current methods for tackling skateboarding with humanoid robots either require human demonstrations or utilise analytic steering models that require system identification. In this paper, we introduce a deep reinforcement learning framework that does not rely on prior data such as demonstrations and enables robots such as the K1 humanoid to learn to skateboard using model-free learning. The method combines task rewards with exploit-blocking terminations. We introduce a number of constraints restricting behaviours such as losing the planted board foot, standing with both feet on the board, excessive twisting, separation from the board, and unstable board motion. Their changing prevalence during training produces an implicit progression from stance control towards sustained robot--board locomotion.

Summaries are written automatically from each paper's text. Figures and videos come from the paper and its project page; all credit goes to the authors.