Learning Diverse Natural Behaviors for Enhancing the Agility of Quadrupedal Robots

TL;DR

This work tackles animal-like agility in quadrupedal robots by proposing an integrated BBC+TSC controller that can learn and deploy a diverse set of natural behaviors. A semi-supervised InfoGAIL framework drives the Basic Behavior Controller to acquire multimodal dog-like styles from motion capture data, with a discrete latent variable $\mathbf{c}$ and a continuous $\epsilon$, while Regularized Information Maximization (RIM) helps manage data imbalance. A privileged-learning based Task-Specific Controller coordinates the BBC using depth imagery, and an Evolutionary Adversarial Simulator Identification (EASI) aligns the simulator with reality to enable efficient sim-to-real transfer. The combination yields autonomous, high-speed quadrupedal performance (average $1.1$ m/s, peak $3.2$ m/s in hurdling) and establishes a new agility benchmark, highlighting practical potential for real-world deployment and human-robot collaboration in dynamic environments.

Abstract

Achieving animal-like agility is a longstanding goal in quadrupedal robotics. While recent studies have successfully demonstrated imitation of specific behaviors, enabling robots to replicate a broader range of natural behaviors in real-world environments remains an open challenge. Here we propose an integrated controller comprising a Basic Behavior Controller (BBC) and a Task-Specific Controller (TSC) which can effectively learn diverse natural quadrupedal behaviors in an enhanced simulator and efficiently transfer them to the real world. Specifically, the BBC is trained using a novel semi-supervised generative adversarial imitation learning algorithm to extract diverse behavioral styles from raw motion capture data of real dogs, enabling smooth behavior transitions by adjusting discrete and continuous latent variable inputs. The TSC, trained via privileged learning with depth images as input, coordinates the BBC to efficiently perform various tasks. Additionally, we employ evolutionary adversarial simulator identification to optimize the simulator, aligning it closely with reality. After training, the robot exhibits diverse natural behaviors, successfully completing the quadrupedal agility challenge at an average speed of 1.1 m/s and achieving a peak speed of 3.2 m/s during hurdling. This work represents a substantial step toward animal-like agility in quadrupedal robots, opening avenues for their deployment in increasingly complex real-world environments.

Figures (11)

• Figure 1: Deployment of the proposed controller.a-h, Environments with various obstacles. i, Wild environment. j-n, Natural behavioral modes.
• Figure 2: Evaluation in the quadrupedal agility challenge.a, The full motion sequence from start to end. b, Command variations during the motion process, with the gray-and-white shaded areas representing different obstacle stages. c, A comparison of our method and the baseline method cheng2024extreme through the bar-jump. d, A comparison of success rates and average speeds between our method and the baseline method in single-obstacle environments. e, Success rates for different obstacle heights. The "Operator" refers to a human remotely controlling the BBC. f, Success rates and average speeds of different methods in the full quadrupedal agility challenge.
• Figure 3: Natural jump for high-speed hurdling.a, Two jumping phases. b, Contact sequence of each leg during hurdling. c, Command and actual linear velocity variations. d, Body height variation.
• Figure 4: Evaluation of the BBC's multimodal behaviors.a, Common dog behaviors. b, Learned multimodal behaviors for the quadrupedal robot. c, Illustration of command following. d, Command following error. e, The impact of latent shifting variable changes. f, The proportions of the 5 behavior modes in motion capture data. g, NMI ($\uparrow$) and ENT ($\downarrow$) of different algorithms. h, Visualization of behavioral features across different algorithms.
• Figure 5: Evaluation of the TSC.a, The privileged information observed by the teacher policy, and the student's observation. b, Comparison between our method and two baseline methods: one without privileged learning and one without BYOL. c, The error between the predicted relative yaw angle and the true value.