Learning vision-based agile flight via differentiable physics

Abstract

Autonomous aerial robot swarms promise transformative applications, from planetary exploration to search and rescue in complex environments. However, navigating these swarms efficiently in unknown and cluttered spaces without bulky sensors, heavy computation or constant communication between robots remains a major research problem. This paper introduces an end-to-end approach that combines deep learning with first-principles physics through differentiable simulation to enable autonomous navigation by several aerial robots through complex environments at high speed. Our approach directly optimizes a neural network control policy by backpropagating loss gradients through the robot simulation using a simple point-mass physics model. Despite this simplicity, our method excels in both multi-agent and single-agent applications. In multi-agent scenarios, our system demonstrates self-organized behaviour, which enables autonomous coordination without communication or centralized planning. In single-agent scenarios, our system achieved a 90% success rate in navigating through complex unknown environments and demonstrated enhanced robustness compared to previous state-of-the-art approaches. Our system can operate without state estimation and adapt to dynamic obstacles. In real-world forest environments, it navigates at speeds of up to 20 m s⁻¹, doubling the speed of previous imitation-learning-based solutions. Notably, all these capabilities are deployed on a budget-friendly US$21 computer, which costs less than 5% of the GPU-equipped board used in existing systems.