Deep Drone Racing: From Simulation to Reality with Domain Randomization

What this paper solves

Learning a drone racing policy directly in the real world is sample-inefficient and dangerous. Simulation training is safe and fast but policies trained in simulation fail to transfer when the visual appearance of the simulated world does not match reality. This paper applies domain randomization to the simulation environment during training, forcing the network to learn representations that are invariant to visual appearance rather than tied to any specific render style.

The result is a policy trained entirely in simulation that transfers zero-shot to physical hardware, racing through gates at speeds up to 8 m/s without any fine-tuning on real data. This paper is the domain randomization baseline for quadrotor sim-to-real transfer.

Key mathematical framework

Domain randomization

Let $\xi \sim p(\xi )$ be a randomization parameter sampled from a distribution over visual domain parameters (gate texture, background texture, lighting, camera noise). The network is trained on observations$o_{\xi }$ drawn from the randomized domain:

The key insight is that a policy trained to minimize loss under the entire distribution $p(\xi )$ must learn features that are invariant across visual domains. If the real world falls within (or close to) the support of $p(\xi )$, the trained policy will generalize.

Gate pose estimation

The input to the network is the estimated gate pose rather than raw pixels. A separate gate detection module estimates the gate center position and orientation in camera frame:

This modular design separates perception (gate detection, domain-dependent) from control (gate-relative navigation, domain-invariant). Domain randomization is applied to the perception module; the control module receives a consistent gate-relative state regardless of visual domain.

Network architecture and training

The policy network maps gate-relative state to body rate commands. Let$s=(p_{gc},v,q)\in {\mathbb{R}}^{10}$ be the state consisting of gate-relative position, velocity, and attitude. The network outputs body rates ${\omega }^B\in {\mathbb{R}}^3$ and thrust:

The training loss combines imitation of a model-based expert and a trajectory smoothness regularizer:

Empirical results

• Zero-shot transfer: Policy trained purely in simulation transfers to physical hardware without real-world fine-tuning, achieving 100% gate completion in nominal conditions
• Speed: Successfully races through gates at up to 8 m/s, competitive with the AlphaPilot system
• Domain gap ablation: Without domain randomization, the policy achieves 0% gate completion on the physical platform despite 100% in simulation
• Compute: Inference on an Intel NUC (no discrete GPU) at 30 Hz

What this means for robotic automation

This paper makes the empirical case for domain randomization as a zero-cost sim-to-real transfer technique: the gate completion rate without randomization is 0% on the physical platform; with randomization it is 100%. The technique adds no real-world data collection, no manual annotation, and no fine-tuning after training. The only cost is simulation diversity, which is effectively free.

The modular perception-control decomposition is architecturally important. Domain randomization is applied to the perception layer (gate detection) while the control layer (gate-relative navigation) is trained in a canonical coordinate frame. This separation is the correct design pattern when perception is domain-dependent but the control problem has a canonical formulation that is domain-invariant.