Phase-Aware Policy Learning for Skateboard Riding of Quadruped Robots via Feature-wise Linear Modulation

Minsung Yoon, Jeil Jeong, Sung‐Eui Yoon

TL;DR

Researchers have developed Phase-Aware Policy Learning (PAPL), a reinforcement learning framework that enables quadruped robots to ride skateboards with human-like agility. By using a "phase clock" and Feature-wise Linear Modulation (FiLM), the robot successfully transitions between kicking for speed and tilting the board to steer, achieving significantly higher energy efficiency than traditional walking.

What Problem Does This Paper Address?

While legged robots are great at navigating rough terrain, they are notoriously energy-inefficient and slow for long-distance travel. One solution is to have them use "personal mobility devices" like skateboards. However, skateboarding is a complex, multi-modal task:

  1. Conflicting Objectives: Pushing (kicking the ground) requires different balance and contact patterns than carving (tilting the deck to steer).
  2. Underactuated Dynamics: The robot and the board are two separate floating bases that only interact through friction and contact.
  3. Perception Gap: To stay on the board without mechanical attachments, the robot needs to precisely sense the board's position relative to its own body in real-time.

Previous attempts often simplified the problem by "gluing" the robot's feet to the board in simulation, which fails to capture the nuances of real-world riding and steering.

Methods

The authors propose PAPL, which treats skateboarding as a cyclic process managed by a Phase Clock.

  • Phase-Conditioned Architecture: Instead of training separate neural networks for pushing and carving, they use Feature-wise Linear Modulation (FiLM). This allows a single "unified" policy to adapt its behavior based on the current phase (Pushing, Transition, or Carving) by applying affine transformations to the network's internal features.
  • Two-Stage Privileged Learning:
    1. Teacher Phase: A "critic" network is trained in simulation with access to "privileged" information (like exact friction coefficients and board velocity).
  • Student Phase (Distillation): Using the DAgger algorithm, the robot learns to estimate those hidden properties using only its onboard sensors—specifically a belly-mounted RGB camera and joint encoders.
  • Exteroceptive Feedback: The robot uses a belly-mounted camera to generate semantic masks of the skateboard deck, allowing it to "see" where the board is and adjust its posture to prevent slipping off.

Datasets

The study primarily used synthetic data generated within the Isaac Gym physics simulator.

  • Training Environment: 4,096 parallel environments featuring a Unitree Go1 quadruped and a standard-sized skateboard.
  • Domain Randomization: To ensure the policy works in the real world, the authors randomized parameters like payload mass, friction, motor strength, and board dimensions during training.
  • Real-World Validation: The final policy was tested on a physical Unitree Go1 robot on various surfaces (sidewalks, low-light areas).

Key Insights and Ideas

  • FiLM vs. Mixture-of-Experts (MoE): The researchers found that FiLM outperformed MoE. While MoE tries to switch between specialized "expert" sub-networks, FiLM allows for smoother transitions and better knowledge sharing across phases, which is vital for maintaining balance during the split-second transitions between kicking and gliding.
  • The "Long Tail" of Power: Power consumption analysis showed that while pushing requires high bursts of energy, the "carving" and "gliding" phases are incredibly cheap. This makes skateboarding a net win for robot battery life over long distances.
  • Visuals are Vital: Ablation studies showed that without the belly-mounted camera (the "No-Extero" variant), the robot frequently lost track of the board and suffered "constraint violations" (falling off or drifting away).

Key Results

  • Efficiency: Skateboarding was found to be significantly more energy-efficient than both standard walking and "wheel-leg" hybrid locomotion. The mean power consumption for skateboarding was 0.1093, compared to 0.2229 for walking.
  • Command Tracking: The PAPL policy achieved a much higher "command area" (the range of speeds and turn rates it can handle reliably) compared to versions without phase-awareness or exteroceptive sensing.
  • Zero-Shot Transfer: The policy trained entirely in simulation worked on a real robot in the real world without any fine-tuning, even when faced with external pushes, uneven sidewalks, and low-light conditions.

Limitations and Future Work

  • Fixed Phase Period: Currently, the "phase clock" (how long the robot spends pushing vs. carving) is a pre-defined parameter. The authors suggest that a high-level navigation policy should eventually control this clock dynamically based on the environment (e.g., pushing harder when going uphill).
  • Navigation: The current system focuses on the mechanics of riding. It doesn't yet "look ahead" to avoid obstacles or plan a route; it simply follows velocity commands from a remote or a higher-level planner. Integrating front-facing cameras for autonomous navigation is the next logical step.

Summary updated

Abstract

Skateboards offer a compact and efficient means of transportation as a type of personal mobility device. However, controlling them with legged robots poses several challenges for policy learning due to perception-driven interactions and multi-modal control objectives across distinct skateboarding phases. To address these challenges, we introduce Phase-Aware Policy Learning (PAPL), a reinforcement-learning framework tailored for skateboarding with quadruped robots. PAPL leverages the cyclic nature of skateboarding by integrating phase-conditioned Feature-wise Linear Modulation layers into actor and critic networks, enabling a unified policy that captures phase-dependent behaviors while sharing robot-specific knowledge across phases. Our evaluations in simulation validate command-tracking accuracy and conduct ablation studies quantifying each component's contribution. We also compare locomotion efficiency against leg and wheel-leg baselines and show real-world transferability.
BibTeX
@misc{yoon2026phase,
  author = {Minsung Yoon and Jeil Jeong and Sung‐Eui Yoon},
  title = {{Phase-Aware Policy Learning for Skateboard Riding of Quadruped Robots via Feature-wise Linear Modulation}},
  year = {2026},
  eprint = {2602.09370},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2602.09370},
}
Figure