Reproduce “D-VAT: End-to-End Visual Active Tracking for Micro Aerial Vehicles” on OpenResearch

About this paper

Alberto Dionigi, Simone Felicioni, Mirko Leomanni, Gabriele Costante

Researchers from the University of Perugia developed D-VAT, an end-to-end visual active tracking method for Micro Aerial Vehicles that directly translates monocular camera input into continuous control commands. The system demonstrated robust, precise, and collision-free tracking in diverse simulated environments and successfully transferred to a real quadrotor without fine-tuning.

Key insights

• End-to-end DRL with continuous control can effectively learn complex 3D MAV tracking policies directly from monocular vision, outperforming both modular and discrete-action DRL baselines by avoiding restrictive assumptions.
• The asymmetric actor-critic framework, combined with a comprehensive reward function, facilitates efficient learning of robust tracking policies without needing explicit knowledge of MAV dynamics or intermediate state estimation.
• Extensive domain randomization during training, including augmented dynamics and parameter variation, is crucial for achieving strong generalization to diverse, visually complex environments and enabling successful zero-shot transfer of policies to real MAV platforms.

Summary by alphaXiv.