Learning a Centroidal Motion Planner for Legged Locomotion (2011.02818v1)

Abstract: Whole-body optimizers have been successful at automatically computing complex dynamic locomotion behaviors. However they are often limited to offline planning as they are computationally too expensive to replan with a high frequency. Simpler models are then typically used for online replanning. In this paper we present a method to generate whole body movements in real-time for locomotion tasks. Our approach consists in learning a centroidal neural network that predicts the desired centroidal motion given the current state of the robot and a desired contact plan. The network is trained using an existing whole body motion optimizer. Our approach enables to learn with few training samples dynamic motions that can be used in a complete whole-body control framework at high frequency, which is usually not attainable with typical full-body optimizers. We demonstrate our method to generate a rich set of walking and jumping motions on a real quadruped robot.

• The paper presents a centroidal neural network that learns fast motion planning for legged locomotion, reducing computational overhead by up to 41x.
• It decomposes kino-dynamic optimization into centroidal and kinematic components, enabling efficient mapping of desired trajectories to full-body motions.
• Real-world experiments on a quadruped robot demonstrate robust walking and jumping at a 100Hz replanning rate, validating the method’s practical applicability.

Overview of "Learning a Centroidal Motion Planner for Legged Locomotion"

The paper "Learning a Centroidal Motion Planner for Legged Locomotion" authored by Julian Viereck and Ludovic Righetti addresses the computational challenges associated with real-time whole-body motion planning in legged robots. The work presents a learning-based method for generating dynamic locomotion tasks in real-time, fundamentally shifting from traditional trajectory optimization methods that are computationally expensive and not amenable to high-frequency replanning.

Core Contributions

The primary contribution of this paper is a machine learning-driven approach to predict centroidal motions for legged robots, which provides significant speed improvements over traditional methods. Specifically, the paper involves:

1. Centroidal Neural Network: The paper introduces a centroidal neural network that predicts desired centroidal trajectories given the robot's current state and desired contact sequences. The network is trained based on data generated from an existing whole-body optimization framework, significantly reducing the need for computationally intensive real-time optimization.
2. Kino-Dynamic Decomposition: By leveraging a kino-dynamic decomposition strategy, the authors separate the problem into centroidal dynamics and kinematic optimizations. The process involves utilizing inverse kinematics to map predicted centroidal trajectories to full-body motions efficiently.
3. Motion Pattern Learning: The research shows advanced learning of motion patterns that enable the generation of viable motion plans that extend beyond the training domain, successfully addressing varying contact sequences required for robust locomotion.
4. Demonstrations on Real Hardware: Experimental validations on a quadrupedal robot show that the approach can generate robust walking and jumping motions at a rate of 100 Hz, one order of magnitude faster than classical trajectory optimization methods.

Detailed Evaluation

The authors conducted thorough experiments on a Solo12 quadruped robot, noting low tracking errors for both center of mass positioning and base orientation during thirty dynamic tasks, including static walks and jumping. The centroidal neural network was shown to achieve timing performance where the prediction time was consistently lower than the motion duration, facilitating real-time application.

The robustness of the approach is highlighted through its application to varied locomotion patterns that were not part of the training dataset, indicating that learned motion patterns can be generalized effectively. Furthermore, the method exhibits computational speedups of 22.1x to 41.3x compared to traditional kino-dynamic optimization, proving its practical value in real-world applications.

Implications and Future Directions

From a practical standpoint, this research significantly impacts the deployment of legged robots in dynamic environments where computationally efficient and quick adaptation to changing conditions are critical. Theoretically, this work advances the integration of machine learning with classical robotics control frameworks, showing potential for further development in both domains.

An intriguing area for future research includes extending the approach to non-flat terrains and more complex robotic structures, such as humanoids. Additionally, there is potential for integrating this method with reinforcement learning to further improve the adaptability and efficiency of legged locomotion.

This paper sets a foundational step towards real-time adaptive locomotion in legged robots and opens exciting avenues for research in combining machine learning with robotic control principles.