Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Learning to Walk and Fly with Adversarial Motion Priors

Giuseppe L'Erario, Drew Hanover, Angel Romero, Yunlong Song, Gabriele Nava, Paolo Maria Viceconte, Daniele Pucci, Davide Scaramuzza

TL;DR

This work tackles multimodal locomotion by enabling autonomous transitions between walking and flying in an aerial humanoid robot via Adversarial Motion Priors (AMP) and reinforcement learning. It learns walking from human-like gaits and flying from trajectory-optimized data, while environment feedback through a terrain-aware reward induces spontaneous mode switching without explicit planners. The approach outperforms baselines by producing smooth, energy-aware transitions and adapting to complex terrains; ablations show that combining walking and flying priors yields the best thrust efficiency. The results, demonstrated in simulation with iRonCub and supported by comparisons to trajectory optimization, suggest AMP’s potential to broaden the capabilities of aerial humanoid robots for tasks such as search, surveillance, and exploration, with future work aiming at real-world validation and additional motion priors.

Abstract

Robot multimodal locomotion encompasses the ability to transition between walking and flying, representing a significant challenge in robotics. This work presents an approach that enables automatic smooth transitions between legged and aerial locomotion. Leveraging the concept of Adversarial Motion Priors, our method allows the robot to imitate motion datasets and accomplish the desired task without the need for complex reward functions. The robot learns walking patterns from human-like gaits and aerial locomotion patterns from motions obtained using trajectory optimization. Through this process, the robot adapts the locomotion scheme based on environmental feedback using reinforcement learning, with the spontaneous emergence of mode-switching behavior. The results highlight the potential for achieving multimodal locomotion in aerial humanoid robotics through automatic control of walking and flying modes, paving the way for applications in diverse domains such as search and rescue, surveillance, and exploration missions. This research contributes to advancing the capabilities of aerial humanoid robots in terms of versatile locomotion in various environments.

Learning to Walk and Fly with Adversarial Motion Priors

TL;DR

This work tackles multimodal locomotion by enabling autonomous transitions between walking and flying in an aerial humanoid robot via Adversarial Motion Priors (AMP) and reinforcement learning. It learns walking from human-like gaits and flying from trajectory-optimized data, while environment feedback through a terrain-aware reward induces spontaneous mode switching without explicit planners. The approach outperforms baselines by producing smooth, energy-aware transitions and adapting to complex terrains; ablations show that combining walking and flying priors yields the best thrust efficiency. The results, demonstrated in simulation with iRonCub and supported by comparisons to trajectory optimization, suggest AMP’s potential to broaden the capabilities of aerial humanoid robots for tasks such as search, surveillance, and exploration, with future work aiming at real-world validation and additional motion priors.

Abstract

Robot multimodal locomotion encompasses the ability to transition between walking and flying, representing a significant challenge in robotics. This work presents an approach that enables automatic smooth transitions between legged and aerial locomotion. Leveraging the concept of Adversarial Motion Priors, our method allows the robot to imitate motion datasets and accomplish the desired task without the need for complex reward functions. The robot learns walking patterns from human-like gaits and aerial locomotion patterns from motions obtained using trajectory optimization. Through this process, the robot adapts the locomotion scheme based on environmental feedback using reinforcement learning, with the spontaneous emergence of mode-switching behavior. The results highlight the potential for achieving multimodal locomotion in aerial humanoid robotics through automatic control of walking and flying modes, paving the way for applications in diverse domains such as search and rescue, surveillance, and exploration missions. This research contributes to advancing the capabilities of aerial humanoid robots in terms of versatile locomotion in various environments.
Paper Structure (25 sections, 17 equations, 8 figures, 4 tables)

This paper contains 25 sections, 17 equations, 8 figures, 4 tables.

Table of Contents

  1. Introduction
  2. Background
  3. Floating-base modeling
  4. Reinforcement Learning
  5. Method
  6. Adversarial Motion Priors
  7. Style-reward
  8. Task-reward
  9. Model representation
  10. Observation Space
  11. Action Space
  12. Discriminator Observation Space
  13. Results
  14. Motion Priors
  15. Experimental Setup
  16. ...and 10 more sections

Figures (8)

  • Figure 1: iRonCub, the aerial humanoid robot, on a complex terrain.
  • Figure 2: The discriminator learns to distinguish between samples from the dataset and samples produced by the agent. The policy $\pi_\sigma$ is trained to imitate the dataset's motion and accomplish a task simultaneously by maximizing the total reward $r(t)$ that expresses the quality of the motion and the task accomplishment.
  • Figure 3: Snapshots of the flying motion obtained using TO.
  • Figure 5: Training curves over 10 runs on the scenario from Sec. \ref{['sec:terrain-aware-locomotion']}.
  • Figure 6: Walk-to-fly maneuver. The robot lands and walks to catch a waypoint on the ground level, takes off, and flies to hit the aerial target.
  • ...and 3 more figures