SKATER: Synthesized Kinematics for Advanced Traversing Efficiency on a Humanoid Robot via Roller Skate Swizzles

The gist: A deep reinforcement learning control framework developed for roller skating on a humanoid robot, SKATER, demonstrates superior locomotion efficiency compared to traditional bipedal walking gaits by significantly reducing joint impact forces and the cost of transport.

Hardware Design

SKATER is a newly developed 25-degree-of-freedom (DoF) humanoid robot equipped with four passive inline wheels at each sole for roller skating The robot has a total mass of 38 kg and a standing height of 140 cm The roller skating foot mechanism represents the core innovation of SKATER Unlike traditional humanoid robots that employ either rigid flat feet or actuated wheels, SKATER uses a specifically optimized roller skating mechanism This optimization is particularly critical during the lateral propulsion phase of roller skating, where the ankle joint must withstand significant eversion and inversion moments

Reinforcement Learning for Humanoid Locomotion

The paper proposes a deep reinforcement learning (DRL) based control framework tailored to the dynamic characteristics of roller skating This framework utilizes an implicit gait reward function designed to guide the agent in learning the swizzle skating gaits Furthermore, a multi-stage curriculum learning strategy is employed to progressively increase task complexity during training The observation space for the Actor network includes the robot’s current state along with states from 4 historical time steps The Critic has access to additional privileged information such as actual base linear velocity and ankle joint orientation angles The action space employs a torque-based PD controller for joint actuation, where the desired position is expressed as pd t = pt + βat

Reward Function Design

The reward function for roller skating differs fundamentally from traditional bipedal walking It is a multiobjective reward function comprising 22 reward terms specifically designed to balance task completion and motion constraints The design does not explicitly introduce kinematic trajectories or gait timing constraints for swizzle Instead, only necessary physical constraints and motion preferences are provided through boundary constraints on inter-foot distance (minimum 0.2 m and maximum 0.5 m) and limb symmetry rewards This setup sees natural emergence of swizzle gaits during the training process

Experimental Validation Results

Simulation analysis revealed a 75.86% reduction in contact force impact intensity Real robot experiments have shown a 63.34% reduction in Cost of Transport (CoT) Peak torques in hip and ankle joints were reduced by approximately 74% and 70%, respectively The robot has achieved a 100% success rate with tests across ground surfaces with various friction coefficients, demonstrating strong control robustness of the learned policy

Conclusion and Future Work

The study concludes that SKATER successfully achieved efficient and smooth swizzle gaits in both simulation and physical experiments, exhibiting significant advantages over traditional bipedal walking gaits Limitations include lateral drift during straight-line skating, lag in velocity tracking response, and the realization of only swizzle gaits with continuous ground contact of both legs Future work can proceed in three directions: investigating high-dynamic gaits such as push steps and single-leg support to enhance maneuverability, integrating perception to enable autonomous navigation in complex environments, and exploring actively switchable passive wheel-foot mechanisms to further improve traversability

--- Page 1 ---

SKATER: Synthesized Kinematics for Advanced Traversing Efficiency on a Humanoid Robot via Roller Skate Swizzles

Junchi Gu, Feiyang Yuan, Weize Shi,Tianchen Huang, Haopeng Zhang, Xiaohu Zhang, Yu Wang, Wei Gao and Shiwu Zhang

Abstract— Although recent years have seen significant progress of humanoid robots in walking and running the frequent foot strikes with ground during these locomotion gaits inevitably generate high instantaneous impact forces which leads to exacerbated joint wear and poor energy utilization. Roller skating, as a sport with substantial biomechanical value, can achieve fast and continuous sliding through rational utilization of body inertia, featuring minimal kinetic energy loss. Therefore, this study proposes a novel humanoid robot with each foot equipped with a row of four passive wheels for roller skating. A deep reinforcement learning control framework is also developed for the swizzle gait with the reward function design based on the intrinsic characteristics of roller skating. The learned policy is first analyzed in simulation and then deployed on the physical robot to demonstrate the smoothness and efficiency of the swizzle gait over traditional bipedal walking gait in terms of Impact Intensity and Cost of Transport during locomotion. A reduction of 75.86% and 63.34% of these two metrics indicate roller skating as a superior locomotion mode for enhanced energy efficiency and joint longevity.

I. INTRODUCTION

Recent advances in locomotion control of legged robots have yielded remarkable movement strategies, including walking [1], running [2], jumping [3] and dancing [4] However, these motions often impose extreme demands on joint actuator torque and velocity undertake significant joint impacts during execution and produce substantial noises. Specifically, in walking and running gaits, robots experience intermittent ground contacts with considerable instantaneous impact forces at each touchdown This not only accelerates joint wear but also substantially reduces energy efficiency. Roller skating, as a low-impact locomotion modality, offers unique biomechanical advantages. Previous study has indicated that roller skating generates approximately 50% of the joint impact forces compared to running [5], with its smooth sliding motion significantly reducing workloads on the knee, hip and ankle joints. Thus, roller skating can leverage wholebody inertia to substantially minimize joint wear and energy dissipation while maintaining comparable traversing capabilities. This characteristic enables roller skating to be an efficient locomotion mode for extended operational endurance and joint longevity of humanoid robots Therefore, this paper presents a newly developed humanoid robot, SKATER (Synthesized Kinematics for Advanced Traversing Efficiency on a Humanoid Robot), that is capable of roller skating with its special foot structure design, as shown in Fig. 1.

Fig. 1: The SKATER system: a humanoid robot equipped with roller skates for learning swizzle locomotion through deep reinforcement learning

SKATER is a newly developed 25-degree-of-freedom (DoF) humanoid robot equipped with four passive inline wheels at each sole for roller skating The robot has a total mass of 38 kg and a standing height of 140 cm

--- Page 2 ---

However, performing roller skating maneuver on humanoid platforms presents numerous control challenges First, roller skating violates the fundamental assumption of static contact between stance foot and ground in traditional bipedal locomotion control [6]. Continuous foot sliding during the stance phase drastically compromises system stability and significantly complicates balance control Second, roller skating involves intricate nonholonomic constraints, where wheels can only roll along specific directions without lateral sliding, introducing additional complexity to motion planning and control <ref:2601.

1. Bold Header: Implicit Gait Reward Function Design

The system can learn stable periodic patterns by designing a reward function grounded in physical consistency, geometric constraints and kinematic symmetry, allowing DRL to spontaneously yield stable periodic patterns without explicit gait timing constraints.

2. Bold Header: Reduced Joint Impact Loads

The improved system can achieve superior joint longevity by reducing peak torques; specifically, the paper demonstrates that the torque reductions are most pronounced in the hip pitch and ankle pitch joints, reaching 74% and 65% − 75%, respectively.

3. Bold Header: Robust Terrain Adaptability

The AI system can maintain stable locomotion across varied surfaces with a 100% success rate on tests involving tile floor, rubber floor, and gravel pavement by adapting skating motions based on friction coefficients.

4. Bold Header: Energy Efficiency Optimization

The robot can achieve enhanced endurance in long-distance tasks by exploiting passive dynamics; the study shows that the swizzle gait achieves a 63.34% reduction in CoT compared to the walking gait, with specific joints like Hip Pitch experiencing a 95.39% reduction.

5. Bold Header: Sim-to-Real Transfer Robustness

The system can be deployed effectively on physical hardware by mitigating discrepancies between simulation and reality using domain randomization techniques and careful calibration of dynamic parameters.

Although recent years have seen significant progress of humanoid robots in walking and running, the frequent foot strikes with ground during these locomotion gaits inevitably generate high instantaneous impact forces, which leads to exacerbated joint wear and poor energy utilization. Roller skating, as a sport with substantial biomechanical value, can achieve fast and continuous sliding through rational utilization of body inertia, featuring minimal kinetic energy loss. Therefore, this study proposes a novel humanoid robot with each foot equipped with a row of four passive wheels for roller skating. A deep reinforcement learning control framework is also developed for the swizzle gait with the reward function design based on the intrinsic characteristics of roller skating. The learned policy is first analyzed in simulation and then deployed on the physical robot to demonstrate the smoothness and efficiency of the swizzle gait over traditional bipedal walking gait in terms of Impact Intensity and Cost of Transport during locomotion. A reduction of 75.86% and 63.34% of these two metrics indicate roller skating as a superior locomotion mode for enhanced energy efficiency and joint longevity.