Accepted to ICRA 2026

Phase-Aware Policy Learning for
Skateboard Riding of Quadruped Robots

via Feature-wise Linear Modulation

Minsung Yoon^* Jeil Jeong^* Sung-Eui Yoon^†

School of Computing, KAIST

^*Equal contribution · ^†Corresponding author

Unitree Go1 riding a skateboard with a belly-mounted RGB camera — Unitree Go1 with a belly-mounted RGB camera for closed-loop skateboard riding.

Abstract

Controlling a skateboard with a legged robot is difficult due to perception-driven contact and multi-modal objectives across riding phases. We propose Phase-Aware Policy Learning (PAPL), an RL framework that modulates actor and critic networks with phase-conditioned FiLM layers, yielding a unified policy that captures phase-specific behavior while sharing robot-level knowledge. PAPL attains accurate command-following in simulation and transfers zero-shot to hardware.

Method

PAPL framework overview — Asymmetric actor–critic trained with PPO, then distilled via DAgger into observation-history estimators for partial-observability deployment.

Phase clock φ_t: Pushing, Transition, Carving.

FiLM-modulated MLP — Per-layer FiLM modulation conditioned on φ_t.

Simulation Experiments

Diverse simulation validation of PAPL — varied commands, rough terrains, and comparisons with other locomotion modalities.

Tracking error heatmaps and ablations — Tracking-error heatmaps and command-area curves. PAPL covers the broadest low-error region; ablating FiLM, exteroception, or privileged learning each degrades performance.

Power consumption — Motor power over a 30 m traversal — skateboarding consumes less energy than legged and wheel–legged baselines.

Real-World Experiments

Real-world skateboarding demonstrations — Zero-shot transfer across perturbations, low light, and uneven sidewalks.

Indoor.

Outdoor.

BibTeX

@misc{yoon2026phaseawarepolicylearningskateboard,
  title         = {Phase-Aware Policy Learning for Skateboard Riding of
                   Quadruped Robots via Feature-wise Linear Modulation},
  author        = {Minsung Yoon and Jeil Jeong and Sung-Eui Yoon},
  year          = {2026},
  eprint        = {2602.09370},
  archivePrefix = {arXiv},
  primaryClass  = {cs.RO},
  url           = {https://arxiv.org/abs/2602.09370}
}