One Net to Rule Them All: Domain Randomization in Quadcopter Racing Across Different Platforms

TL;DR

This work addresses the challenge of cross-platform generalization in high-speed quadcopter racing by introducing a neural network controller trained with domain randomization (DR) that operates end-to-end from state to motor commands. The methodology defines a parametric quadcopter model and a PPO-based RL framework, with both a general policy and platform-specific fine-tuned variants trained under varied DR levels. Key findings show that the general DR-trained policy transfers to both 3-inch and 5-inch quadcopters, achieving real-world flight on each platform, albeit with a modest speed penalty compared to fine-tuned controllers; increasing DR improves robustness and sim-to-real transfer, while excessive DR degrades speed. The results highlight DR as a viable path toward universal AI controllers for heterogeneous platforms, while also pointing to reward shaping and training duration as important factors for closer-to-time-optimal performance in real-world racing scenarios.

Abstract

In high-speed quadcopter racing, finding a single controller that works well across different platforms remains challenging. This work presents the first neural network controller for drone racing that generalizes across physically distinct quadcopters. We demonstrate that a single network, trained with domain randomization, can robustly control various types of quadcopters. The network relies solely on the current state to directly compute motor commands. The effectiveness of this generalized controller is validated through real-world tests on two substantially different crafts (3-inch and 5-inch race quadcopters). We further compare the performance of this generalized controller with controllers specifically trained for the 3-inch and 5-inch drone, using their identified model parameters with varying levels of domain randomization (0%, 10%, 20%, 30%). While the generalized controller shows slightly slower speeds compared to the fine-tuned models, it excels in adaptability across different platforms. Our results show that no randomization fails sim-to-real transfer while increasing randomization improves robustness but reduces speed. Despite this trade-off, our findings highlight the potential of domain randomization for generalizing controllers, paving the way for universal AI controllers that can adapt to any platform.

Figures (7)

• Figure 1: We train a single neural network to control physically distinct drones. The time-lapse image demonstrates the successful sim-to-real transfer of the reinforcement-learned network, enabling it to perform drone racing with both 3-inch and 5-inch quadcopters.
• Figure 2: This picture shows the drone used for testing. Instead of bolts, four cornerpieces of Moongel are placed on top of the base frame, on which a 3D printed casing with the flight controller is placed. A second layer of Moongel is then placed on the casing, which is then slightly compressed downward by means of two zip-ties.
• Figure 3: Comparison of network sizes: The 64,64,64 architecture achieved the highest mean episode reward among the evaluated architectures, based on three independent runs with 100 million time steps each.
• Figure 4: Comparison of training performance with various input modifications to the 64,64,64 ReLU network.
• Figure 5: Comparison of real-world and simulated flight trajectories for 3-inch and 5-inch quadcopters, using the general policy (for both platforms) and fine-tuned policies (trained on their respective parametric models) with 30%, 20%, 10%, and 0% randomization