Reference-Free Humanoid Skateboarding via Termination-Shaped Reinforcement Learning

Ganeshvar Soenessardien, Michael Przystupa, Kevin Sebastian LuckarXiv 2026 · arXiv

TL;DR

Researchers have successfully trained a humanoid robot to skateboard from scratch using reinforcement learning without any human demonstrations or motion priors. By using "termination-shaped" constraints as an implicit curriculum, the robot independently discovers how to maintain a stable stance and generate forward propulsion on a passive rolling board.

What Problem Does This Paper Address?

Humanoid locomotion has seen massive leaps in walking and running, but "object-mediated locomotion"—where a robot must coordinate its own body with a passive, movable external object—remains a significant hurdle. Skateboarding is a prime example: it requires simultaneous interaction with the ground (to push) and a rolling platform (to ride), all while managing complex contact dynamics and momentum.

Most existing approaches to this problem rely on Adversarial Motion Priors (AMP)—essentially showing the robot human "reference" data so it knows what skating should look like—or complex analytical steering models. This paper asks a fundamental question: Can recognizable skateboarding behavior emerge purely from task rewards and physical constraints, without any human "cheat sheet"?

Methods

The authors utilized Deep Reinforcement Learning (PPO) within the MuJoCo physics engine, specifically using the mjlab framework for massive parallelization. The robot of choice was the Booster K1 humanoid.

The core innovation is Termination-Shaped Reinforcement Learning. Instead of relying solely on a complex reward function, the researchers defined a suite of "exploit-blocking" termination conditions. These act as hard constraints that end an episode if the robot performs an "illegal" move, such as:

  • board_foot_displaced: Terminating if the foot on the board moves more than 5cm from its initial position.
  • robot_board_yaw_misaligned: Terminating if the robot twists too far relative to the board.
  • both_feet_on_board: Preventing the robot from simply standing still on the deck.

By combining these terminations with simple rewards for forward velocity and trunk height, the system creates an implicit curriculum. The robot must first master "stance control" (keeping its foot planted) before it can survive long enough to earn rewards for "locomotion" (moving forward).

Fig. 1: This paper investigates the problem of learning to skateboard on mobile platforms (above: front- and back- view) in humanoids (K1 Booster) with model-free reinforce- ment learning.

Datasets

Not discussed in the paper. The study relies entirely on simulated environments and does not use external motion-capture or trajectory datasets.

Key Insights and Ideas

  • Negative Constraints as a Teacher: The most striking finding is how the termination conditions drive the learning process. As shown in Figure 5, early training is dominated by failures in stance (foot displacement). Once the robot learns to keep its foot still, the failure mode shifts toward locomotion limits (going out of bounds). This "failure-mode curriculum" guides the agent toward the intended mechanics without manual tuning of reward schedules.
  • Function vs. Style: The robot successfully learned the function of skateboarding (propelling itself forward) but developed a distinct, "conservative" style. Instead of a fluid, high-clearance human-like leg swing, it developed a low-clearance shuffle. This suggests that while task rewards are enough for functional success, motion priors are likely necessary to achieve human-like aesthetics.
  • The XOR Stance: The policy consistently converged to what the authors call the "XOR stance"—one foot on the board, one foot off. This is a fundamental requirement for propulsion that the robot discovered entirely on its own.
The distribution of failure modes shifts over time, creating an implicit curriculum.

Key Results

  • Stable Locomotion: The robot achieved sustained skateboarding for hundreds of steps, reaching a mean board speed of 0.27 m/s.
  • High Fidelity to Task: The robot maintained the correct one-foot-on-board stance (XOR-stance rate) 97.6% of the time.
  • Efficiency: The entire behavior was learned in roughly 28.7 hours on a single consumer workstation (NVIDIA RTX 3070 Ti), demonstrating that complex object-mediated skills don't necessarily require supercomputing clusters.
  • Emergent Stepping: Qualitative analysis (Figure 6) shows a clear propulsive cycle: load/contact, early push, toe-off, foot lift, and repositioning.
Side view of the emergent propulsive stepping sequence.

Limitations and Future Work

  • Conservative Movement: The "near-ground" reward term, while helpful for stability, likely discouraged the robot from developing a more athletic, high-swinging push.
  • Simplified Board Dynamics: The board used was a caster-wheeled platform rather than a traditional skateboard with "trucks." Consequently, the robot could not learn to steer by leaning (lean-to-steer).
  • Straight-Line Only: The current work focuses on forward propulsion; steering and complex maneuvers remain for future research.
  • Sim-to-Real: While the authors used domain randomization (varying mass, friction, etc.) to prepare for the real world, the controller has not yet been validated on physical hardware.

Summary updated

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.
BibTeX
@misc{soenessardien2026reference,
  author = {Ganeshvar Soenessardien and Michael Przystupa and Kevin Sebastian Luck},
  title = {{Reference-Free Humanoid Skateboarding via Termination-Shaped Reinforcement Learning}},
  year = {2026},
  eprint = {2611.26054},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.26054},
}
Figure