Percorrer por data de Publicação, começado por "2026-04-22"
A mostrar 1 - 1 de 1
Resultados por página
Opções de ordenação
- VIRIATO: Visual-Action Reinforcement Integrator for Actor-Critic with Temporal ObservationsPublication . Campanhã, João; Neves, Francisco; Pinto, Andry; BENEDITA CAMPOS NEVES MALHEIRO, MARIA; Fonseca, Pedro; Moreira, António; Neto, Pedro; Morais, Pedro; Lima, JoséThe Visual-Action Reinforcement Integrator for Actor-Critic with Temporal Observations (VIRIATO) is a compact multi-input feature-extractor architecture designed to enable robust visual navigation of Unmanned Aerial Vehicles (UAVs) conducting close-range inspection of photovoltaic arrays. The target task of low-altitude flight over dynamic, visually variable surfaces without privileged information is inherently partially observable. VIRIATO augments stacked image observations with a short history of recent past actions, producing a richer latent state for the Soft Actor-Critic (SAC) agent. Training is performed with domain randomization to expose the policy to diverse lighting, backgrounds and panel layouts. In simulation, VIRIATO yields faster learning and improved sample efficiency compared to a standard image-only Convolutional Neural Network (CNN) feature extractor, achieving lower position and yaw errors and substantially better robustness under image perturbations while retaining high task completion rates. The architecture is intentionally simple and general: it improves temporal awareness without adding complex recurrence, and it could be adapted to other perception-driven robotic tasks. These results demonstrate that integrating historical action data with visual encoding, together with domain randomization, is an effective way to achieve reliable autonomous vision-based navigation.