RAPTOR: A foundation policy for quadrotor control.

Area of Science:

• Robotics
• Artificial Intelligence
• Control Systems

Background:

• Modern robotic control systems, particularly those using reinforcement learning (RL), often overfit to specific environments.
• This specialization leads to breakdowns when encountering minor variations, such as the simulation-to-reality gap, necessitating extensive retraining.
• Quadrotor control systems typically require system identification and re-training for even small changes.

Purpose of the Study:

• To present RAPTOR, a method for training a highly adaptive foundation policy for quadrotor control.
• To enable a single neural network policy to control a wide variety of quadrotors.
• To achieve rapid, zero-shot adaptation to unseen quadrotor platforms.

Main Methods:

• Trained a single, end-to-end neural network policy using a meta-imitation learning algorithm.
• Sampled 1000 diverse quadrotors, training a teacher policy for each using RL.
• Distilled 1000 teacher policies into a single adaptive student policy with a recurrent hidden layer for in-context learning.

Main Results:

• A compact, three-layer policy with 2084 parameters demonstrated sufficient capability for zero-shot adaptation.
• The foundation policy successfully controlled 10 different real quadrotors with varying characteristics (weight, motors, frames, propellers, flight controllers).
• The policy adapted to unseen quadrotors within milliseconds and performed robustly under various conditions, including trajectory tracking, indoor/outdoor flight, wind, and physical interaction.

Conclusions:

• RAPTOR enables the creation of a single, highly adaptive foundation policy for quadrotor control.
• This approach significantly enhances robotic system adaptability, overcoming the limitations of environment-specific training.
• The method facilitates rapid, zero-shot adaptation to novel platforms and conditions, paving the way for more versatile robotic applications.