Towards Dynamic Quadrupedal Gaits: A Symmetry-Guided RL Hierarchy Enables Free Gait Transitions at Varying Speeds

Jiayu Ding, Xulin Chen, Garrett E. Katz, Zhenyu Gan

TL;DR

This paper introduces a novel reinforcement learning framework that uses symmetries inherent in quadrupedal locomotion to enable a single policy to generate a wide range of gaits (trotting, bounding, galloping, etc.) and smoothly transition between them at varying speeds. This approach significantly simplifies gait control by reducing the need for manual tuning and predefined trajectories, leading to more adaptable and efficient robot locomotion.

What Problem Does This Paper Address?

Quadrupedal robots are promising for various applications, but controlling their diverse gaits and enabling smooth transitions between them remains a significant challenge. Existing methods often rely on laborious manual tuning of numerous parameters, predefined footfall sequences, or complex trajectory optimization. While reinforcement learning (RL) offers a powerful alternative, it typically requires extensive reward function engineering. This paper addresses the need for a unified, adaptive, and efficient control framework that can generate a spectrum of dynamic quadrupedal gaits and transitions without explicit, hand-coded gait definitions.

Methods

The core of this work is a symmetry-guided reinforcement learning approach. The authors leverage the intrinsic temporal, morphological, and time-reversal symmetries found in quadrupedal locomotion.

  • Symmetry-Guided Reward Function: Instead of relying on predefined gaits, the reward function is designed to explicitly incorporate these symmetries. This inductive bias guides the RL agent to learn policies that respect these natural properties, simplifying the learning process and reducing the dimensionality of the search space.
  • Unified Policy: A single policy network (an MLP) is trained to generate all the desired gaits. This is achieved by mapping user commands (desired speed and gait type) into a set of gait parameters (leg phase offsets and duty factors) that are then fed into the policy.
  • Dynamic Gait Parameterization: The stride period (T) and duty factor (β) are dynamically parameterized based on the commanded velocity, allowing for smooth scaling of gait characteristics with speed. This is derived from insights into passive dynamics and empirical calibration.
  • Time-Reversal Mapping: To ensure consistent backward locomotion, a phase mapping is introduced that effectively mirrors the gait when the commanded velocity is negative.
  • Reinforcement Learning Algorithm: Proximal Policy Optimization (PPO) is used for training the policy in simulation.

Datasets

The paper does not explicitly mention using pre-existing datasets for training. Instead, the training is performed in simulation using the Isaac Gym environment, which provides 1024 parallel environments. Domain randomization and velocity/gait resampling are employed during training to enhance robustness and generalization. The evaluation is conducted both in simulation and on a real Unitree Go2 robot.

Key Insights and Ideas

  • Symmetry as a Powerful Inductive Bias: The paper strongly demonstrates that incorporating natural symmetries into the RL reward function is a highly effective way to simplify complex control tasks like quadrupedal gait generation. This reduces the need for manual engineering and leads to more generalizable policies.
  • Unified Control for Diverse Gaits: A single policy can learn to generate a wide range of gaits (trotting, bounding, half-bounding, galloping) and smoothly transition between them, which is a significant step towards more versatile robotic locomotion.
  • Dynamic Gait Parameterization is Crucial: Linking gait parameters like stride period and duty factor to commanded velocity is essential for achieving smooth and efficient locomotion across different speeds.
  • Time-Reversal Symmetry for Directional Robustness: Explicitly enforcing time-reversal symmetry is key to ensuring that backward locomotion is as robust and efficient as forward locomotion.
  • Morphological Symmetry for Efficiency: Morphological symmetry appears to be the primary driver of energetic efficiency in locomotion.

Key Results

  • Robust Velocity Tracking: The learned policy achieved accurate velocity tracking across various commanded speeds and gaits, converging within approximately 1 second.
  • Accurate Gait Tracking and Transitions: The robot successfully reproduced desired footfall patterns and executed smooth transitions between gaits in both simulation and hardware. Gait consistency errors were low (GC = 0.07).
  • Improved Energy Efficiency: The symmetry-guided policy demonstrated a lower Cost of Transport (CoT) compared to policies trained without morphological or time-reversal symmetry, especially at positive velocities. The no-morphology variant was 30-50% more costly.
  • Hardware Validation: The policy was successfully deployed on a Unitree Go2 robot, showing stable locomotion and reliable tracking even on uneven terrain, demonstrating good sim-to-real transfer.

Limitations and Future Work

  • Limitations:
    • The paper focuses on a specific set of gaits. Extending to more complex gaits or highly dynamic maneuvers might require further research.
  • While robustness to uneven terrain was shown, the extent of its generalization to highly unstructured or unpredictable environments is not fully explored.
  • The paper acknowledges that the time-reversal symmetry enforcement primarily enhances directional robustness, suggesting that further work might be needed to optimize backward gaits beyond just mirroring forward ones.
  • Future Work:
    • The authors suggest that the framework could be extended to more complex locomotion tasks, such as manipulation while walking or navigating highly dynamic environments.
  • Further exploration of the interplay between different symmetries and their impact on specific gait characteristics could lead to even more refined control.
  • Investigating how to further reduce the energy cost of locomotion, potentially by exploring more advanced symmetry representations or optimization techniques, is a promising direction.

Summary updated

Abstract

Quadrupedal robots exhibit a wide range of viable gaits, but generating specific footfall sequences often requires laborious expert tuning of numerous variables, such as touch-down and lift-off events and holonomic constraints for each leg. This paper presents a unified reinforcement learning framework for generating versatile quadrupedal gaits by leveraging the intrinsic symmetries of dynamic legged systems. We propose a symmetry-guided reward function design that incorporates temporal, morphological, and time-reversal symmetries, streamlining gait design, accelerating learning, and enhancing robot agility. By focusing on preserved symmetries and natural dynamics, our approach eliminates the need for predefined trajectories or footfall sequences, enabling smooth transitions between diverse locomotion patterns such as trotting, bounding, half-bounding, and galloping. Implemented on the Petoi Bittle robot, our method demonstrates robust performance across a range of speeds in both simulations and hardware tests, significantly improving gait adaptability without extensive reward tuning or explicit foot placement control. This work provides insights into dynamic locomotion strategies and underscores the crucial role of symmetries in robotic gait design.
BibTeX
@misc{ding2025towards,
  author = {Jiayu Ding and Xulin Chen and Garrett E. Katz and Zhenyu Gan},
  title = {{Towards Dynamic Quadrupedal Gaits: A Symmetry-Guided RL Hierarchy Enables Free Gait Transitions at Varying Speeds}},
  year = {2025},
  eprint = {2510.10455},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2510.10455},
}
Figure