Energy-Efficient Omnidirectional Locomotion for Wheeled Quadrupeds via Predictive Energy-Aware Nominal Gait Selection

TL;DR

The paper tackles energy-aware omnidirectional locomotion for wheeled quadrupeds by introducing a hierarchical control framework that first predicts energy consumption over a $1\,s$ horizon to select an energy-efficient nominal gait, then applies a residual RL policy to refine actions. It introduces an augmented gait library that blends wheel and leg motions and couples it with a power-prediction module and a residual RL controller. The approach is validated in both simulation (IsaacGym) and real-world experiments on a modified Unitree Go1, achieving up to $35\%$ energy reductions while maintaining velocity tracking and demonstrating robustness to disturbances. This work advances practical, energy-efficient control for hybrid wheeled-legged robots, enabling longer operation and better adaptability in diverse environments.

Abstract

Wheeled-legged robots combine the efficiency of wheels with the versatility of legs, but face significant energy optimization challenges when navigating diverse environments. In this work, we present a hierarchical control framework that integrates predictive power modeling with residual reinforcement learning to optimize omnidirectional locomotion efficiency for wheeled quadrupedal robots. Our approach employs a novel power prediction network that forecasts energy consumption across different gait patterns over a 1-second horizon, enabling intelligent selection of the most energy-efficient nominal gait. A reinforcement learning policy then generates residual adjustments to this nominal gait, fine-tuning the robot's actions to balance energy efficiency with performance objectives. Comparative analysis shows our method reduces energy consumption by up to 35\% compared to fixed-gait approaches while maintaining comparable velocity tracking performance. We validate our framework through extensive simulations and real-world experiments on a modified Unitree Go1 platform, demonstrating robust performance even under external disturbances. Videos and implementation details are available at \href{https://sites.google.com/view/switching-wpg}{https://sites.google.com/view/switching-wpg}.

Figures (6)

• Figure 1: Real-world experiment on omnidirectional locomotion: forward, turning and lateral movement of our wheeled quadrupedal robot.
• Figure 2: Our low-cost wheeled-legged robot and the exploded view of the modified structure. We have open-sourced our hardware design and firmware code, which is available for download from the project site.
• Figure 3: Overview of the proposed framework: By predicting energy consumption for each gait, the augmented nominal gait with highest energy efficiency is selected from the gait library, generating both wheel velocities and desired end-effector positions for the legs. A policy network, trained in simulation, compensates for the limitations of the nominal gait. It takes as input the history of actions, estimated robot states $z_t$, user commands $v_t^{\text{cmd}}$, proprioceptive measurements, and augmented nominal gait information $p_t$. The network then outputs residuals for the desired end-effector positions of the legs, wheel velocities, and phase adjustments. Together, the nominal gait and the learned residual policy determine the robot's movements.
• Figure 4: Top view of our wheeled quadrupedal robot. The commanded velocity $v_x^{\text{cmd}}, v_y^{\text{cmd}}$ and $\omega^{\text{cmd}}$ is shown in the robot's base frame.
• Figure 5: Commanded velocity tracking error and estimated power for three tasks. We respectively sample $v_x^{\text{cmd}}\in[-1,1]$m/s, $v_y^{\text{cmd}}\in[-0.7, 0.7]$m/s and $\omega^{\text{cmd}}\in[-0.7, 0.7]$rad/s for each task.