Self-Predictive Representation for Autonomous UAV Object-Goal Navigation

Researchers Angel Ayala, Donling Sui, Francisco Cruz, Mitchell Torok, Mohammad Deghat, and Bruno J. T. Fernandes from UNSW Sydney and other institutions have introduced AmelPredSto, a novel self-predictive representation model designed to enhance data sample efficiency in autonomous UAV object-goal navigation (OGN). Published on arXiv on April 22, 2026, this work addresses a critical bottleneck in reinforcement learning (RL) for drones: the need for massive amounts of data to learn effective navigation policies. Traditional approaches rely on relative or absolute coordinates to move from point A to point B, rather than directly searching for a target object. AmelPredSto changes that by learning compressed, predictive state representations from raw sensory inputs, enabling drones to infer the location of unknown objects without explicit coordinate guidance.

The team formalized the unknown target location problem as a Markov decision process (MDP) and tested various state representation learning (SRL) methods combined with model-free actor-critic RL algorithms. Empirical results show that AmelPredSto, the stochastic variant of their self-predictive model, achieved the best performance, yielding a substantial improvement in RL sample efficiency—up to 40% faster convergence compared to baselines. This means drones can learn to navigate to objects with significantly fewer flight trials, which is critical for real-world applications like search and rescue, precision agriculture, and autonomous delivery. The paper is currently under review for IEEE Transactions on Robotics (T-RO), signaling its potential impact on the field.