- The paper demonstrates the first real-world edge-driven skating by a humanoid robot on unmodified passive inline skates, using proprioceptive reinforcement learning without motion data, imitation learning, or precomputed trajectories.
- The paper combines spherical and ellipsoidal wheel-contact models, asymmetric PPO, a rolling-specific reward, and a success-based command curriculum to overcome contact artifacts and train across roughly 3 billion simulation steps.
- The paper achieves zero-shot hardware deployment with speeds up to 1.8 m/s, dynamic turning, disturbance recovery, and a 50% lower cost of transport than walking at 1 m/s, while retaining limitations in tight turns and dominant-leg gait symmetry.
This paper presents an RL-based control framework that enables the Booster T1 humanoid robot to locomote on completely passive, unmodified consumer inline skates, with a zero-shot sim-to-real deployment demonstrating emergent stroke-and-glide propulsion, dynamic turning, and disturbance rejection (2606.31807). To the authors' knowledge, this is the first real-world demonstration of a bipedal humanoid executing edge-driven propulsion on passive inline skates without human motion data, imitation learning, or precomputed trajectories.
Motivation and problem setting
Passive inline skating poses a fundamentally different control problem from actuated wheeled-legged locomotion. Propulsion cannot be applied as motor torque at the wheel axles; instead, the controller must command precise 6-DoF skate end-effector poses, engaging wheel edges to generate the lateral ground reaction forces needed for forward motion. The co-linear wheel arrangement reduces each foot's support polygon to a narrow line, and the absence of wheel encoders introduces partial observability: wheel speed and slippage must be inferred from proprioception alone. Inline skates were chosen over quad skates for their higher achievable velocities and the absence of truck-based coupling between weight distribution and turn radius, at the cost of a more difficult control problem. The authors contrast their approach with the concurrent SKATER robot (Gu et al., 8 Jan 2026), which uses a custom 25-DoF platform and reward shaping that enforces a continuous-contact swizzle gait; this work instead uses off-the-shelf skates and relaxes continuous-contact constraints entirely.
Hardware integration
The platform is the 23-DoF Booster T1, whose 6-DoF legs (3-DoF hips, 1-DoF knees, 2-DoF ankles) provide the full pose control required for edge engagement. Standard feet are replaced with children's consumer inline skates attached via custom stereolithography mounts printed in Formlabs Tough 1500 resin, chosen for impact resilience. Bearings are upgraded (ILQ-9 Twincam Pro) and wheels substituted with less compressible Hydrogen Street 60/92A wheels to better match the rigid-body assumptions of simulation. The policy controls 21 joints, ignoring the head actuators while retaining arm control for balance recovery. Observations are strictly proprioceptive: dual joint encoders and a torso IMU.
Training pipeline
Training is built on Booster Gym (Wang et al., 18 Jun 2025) in Isaac Gym, using asymmetric actor-critic PPO. The actor receives a 74-dimensional noisy proprioceptive observation; the critic receives an 88-dimensional privileged state including ground-truth quantities, base mass and CoM offset, base velocity and height, and external forces. Actions are 21 joint position targets tracked by a high-rate PD controller.
A central contribution is the treatment of physics-engine contact artifacts. High-resolution meshes produce a "sticky contact patch" that artificially damps rolling; cylindrical primitives clip at edges, causing PhysX to inject artificial restitution impulses that the RL agent readily exploits as free propulsion. Capsule primitives cannot represent a thin wheel profile without degenerating to a sphere or introducing an unrealistic linear contact patch. The authors therefore decouple geometry across the pipeline: spherical primitives fitted to the wheel's bottom curvature are used for training in Isaac Gym, while ellipsoidal primitives are used for sim-to-sim validation in MuJoCo. A policy robust across both abstractions is taken as evidence of generalization to the true physical geometry. Notably, the authors concede that this is a workaround rather than a solution—narrow-profile passive wheel contact remains a persistent challenge even in newer solvers such as Isaac Lab's.
Reward structure
Standard bipedal rewards alone yield stepping or shuffling rather than gliding, so a kinematic rolling reward is added. The average linear velocity of each foot is computed from contact-gated wheel angular velocities; the maximum across feet is compared against the base's forward velocity via a Gaussian kernel. Regularization terms penalize torques, velocities, accelerations, and orientation errors. The full reward table includes a wheel-spin term weighted at 6.0, the highest of the task rewards, alongside a 3.0-weighted feet-swing term.
Success-based command curriculum
Because initial exploration on passive wheels is dominated by terminal falls, velocity commands are discretized into a 2D grid over forward and yaw velocity (lateral velocity fixed at zero, reflecting human skating kinematics), starting from a degenerate distribution at zero velocity. A bin is considered solved when episode-average velocity tracking exceeds a threshold over a minimum duration, at which point adjacent higher-velocity bins are unlocked. Training required roughly 100 million steps to learn standing balance and approximately 3 billion steps (20,000 PPO iterations, about 1.9 years of simulated experience) across 8,092 parallel environments to populate the full command range.
Results
Sim-to-sim ablation
The ablation over 100 MuJoCo trials (commands of 1.0–2.0 m/s forward, ±0.5 rad/s yaw, 60 s episodes) substantiates each component. A policy trained on cylindrical primitives achieves high simulated reward purely by exploiting solver artifacts and transfers with almost no velocity tracking (vx​ error of 1.34 ± 0.33 m/s versus 0.43 ± 0.26 m/s for the deployed policy). Removing the rolling reward produces shuffling gaits with falls and a vx​ tracking error of 1.47 ± 0.28 m/s; removing the curriculum degrades performance less severely (0.70 ± 0.21 m/s) but yields unstable techniques the authors judge too risky for hardware deployment.
Efficiency
The Cost of Transport (CoT) comparison against a Booster Gym walking policy shows that walking is more efficient only below roughly 0.5 m/s commanded velocity. At vxcmd​=1 m/s, the skating policy achieves a CoT of 0.326 versus 0.665 for walking—a 50% reduction—while the walking policy's CoT collapses at higher speeds where skating maintains roughly constant efficiency. Real-world measurements on padded mats (an out-of-distribution, cushioned surface) confirm skating becomes more efficient past approximately 0.6 m/s and sustains higher forward velocities than walking can achieve. The CoT maps also reveal a limitation: the skating policy covers a smaller command range than walking, struggles with tight on-the-spot turning relative to its forward performance, and shows asymmetry along the yaw axis attributable to its reliance on a dominant propulsion leg.
Hardware deployment
The policy transfers zero-shot to the untethered Booster T1, demonstrating acceleration to 1.8 m/s and back to rest, dynamic turning at speed, zero-radius turns, recovery from sustained external pushes, and robustness across wooden flooring, brick paving, and padded mats. Joint trajectory comparisons between Isaac Gym and hardware show discrepancies the authors attribute to surface unevenness and friction mismatch; because skating involves continuous ground contact rather than discrete steps, these mismatches compound and affect the policy more than standard locomotion—an honest caveat on the transfer quality.
Limitations and open questions
The paper is explicit about three limitations. First, physical test-space constraints prevented validating the full velocity command range and tracking accuracy on hardware. Second, despite symmetry-encouraging constraints, the optimization landscape converges on a dominant-leg propulsion strategy—a local minimum that improves survivability by minimizing weight-transfer instability but departs from the alternating high-power strokes of human skaters, suggesting the learned gait remains suboptimal in energy efficiency. Third, the wheel-geometry decoupling is an empirical safeguard against contact artifacts rather than a principled contact model, and its success criterion (robustness across two abstractions) is heuristic. Open questions include how to explicitly incentivize alternating-leg propulsion in the reward structure, how the framework extends to ice skating with its phase-change friction dynamics, and how to build a controller that autonomously transitions between walking and skating modes on a modular skate platform.
Conclusion
This work demonstrates that a general-purpose humanoid with off-the-shelf passive inline skates can learn dynamic, edge-driven skating entirely from reward design, achieving a 50% CoT reduction over walking at moderate speeds and robust zero-shot hardware deployment. Its principal methodological contributions—decoupled spherical/ellipsoidal wheel modeling and a success-based command curriculum for an inherently unstable contact mode—offer a reusable recipe for RL control of underactuated rolling contact, while the dominant-leg gait and unvalidated command range mark the clear boundaries of what has been established.
