HUSKY: Humanoid Skateboarding System via Physics-Aware Whole-Body Control

Venue: RSS 2026 (Sydney, Jul 13–17) · Session: Humanoids · paper #19 Authors: Jinrui Han, Dewei Wang, Chenyun Zhang, Xinzhe Liu, Ping Luo, Chenjia Bai, Xuelong Li arXiv: 2602.03205 · program page

Summary compiled from the arXiv paper (v2); all numbers quoted from the paper. Trend context: RSS 2026 survey.

Figure 1 overview: (a) the Unitree G1 performing complete real-world skateboarding — pushing, steering, and phase transitions; (b) generalization to diverse outdoor scenarios and different skateboards; (c) indoor skateboarding; (d) lean-to-steer behavior achieved by tilting the robot body; (e) robustness to external disturbances (box drop, pole push).

Problem

Humanoid whole-body control frameworks mostly assume static environments; skateboarding forces the robot to control its own moving, underactuated support base with non-holonomic wheel constraints, hybrid contact phases, and a small support polygon at 0.6–0.8 m CoM height. Compared with quadruped skateboarding (12 DoF, 2–4 contacts), the humanoid setting (23 DoF, 1–2 contacts, side-on steering, leg reorientation) is far more unstable.

Method

HUSKY first models the humanoid–skateboard system: a kinematic truck model gives the lean-to-steer equality constraint tan σ = tan λ sin γ (board tilt γ → truck steering σ), combined with a bicycle-model yaw approximation to derive a kinematics-guided tilt reference. The task is formulated as a hybrid dynamical system with pushing and steering phases and phase-specific rewards, trained with PPO (asymmetric actor–critic, 4,096 parallel environments, MuJoCo + IsaacLab API). Three key components: (1) AMP-based style rewards from human pushing motions for natural propulsion; (2) physics-guided tilt/heading rewards for steering; (3) a trajectory-guided transition mechanism using Bezier-planned key-body reference states to bridge pushing↔steering. Sim-to-real relies on analytical identification of the skateboard truck's spring–damper tilt dynamics (from free-decay roll response) plus domain randomization; deployment runs at 50 Hz on the 23-DoF Unitree G1.

Results

In simulation (1,000 episodes, 5 seeds), HUSKY reaches 100.0% success versus 11.1% for a tracking-based pushing baseline and 82.4% for a gait-based one, with the lowest velocity error (0.056), heading error (0.208), and contact error (0.001); transition-mechanism ablations (AMP Transition 85.1%, Translation-only 89.6%, Mixed Initialization 86.0%) confirm the trajectory-guided transitions are essential. Real-world G1 experiments show complete pushing–steering–transition skateboarding indoors and outdoors, on two boards with different tilt stiffness, and under external disturbances; swapping the identified board parameters between boards causes mounting failure or over-leaning, validating the system identification.

Significance

A demonstration that explicit physics modeling (lean-to-steer coupling, identified board compliance) folded into RL rewards can conquer a tightly coupled, dynamically unstable human–object system on real hardware. Relevant to the whole-body-control thread in Review-Humanoid-VLA.

← Back to RSS 2026 survey · RSS-2026-Papers · Home