Robot learning

Reinforcement learning and world models for manipulation, navigation and humanoid control.

Learning-based control lets robots handle tasks that are hard to model by hand. Current student projects use reinforcement learning to pick up objects floating in liquids with a robot arm (starting with reward shaping for the reach phase), reinforcement learning for drone navigation during exploration, and world models for controlling humanoid robots.

Earlier, as part of the CNPq project Apprentice: reinforcement learning for collaborative tasks in multi-agent systems (2019–2023), we used reinforcement learning with mobile robots to locate gas sources in mining environments.

Publications

    1. CROS
      Reward Shaping for Learning the Pre-Skimming Reach Phase with a Robotic Manipulator
      Rocha, Thiago Lages, Azpúrua, Héctor and Macharet, Douglas
      In Brazilian Conference on Robotics (CROS 2026)
    2. SBAI
      Localização de Fontes de Gás Usando Aprendizagem por Reforço com Robôs Móveis na Indústria da Mineração
      Sathler, Mauricio Souza, Cid, André Maciel, Oliveira, Felipe, Freitas, Gustavo, Pessin, Gustavo and Azpúrua, Héctor Sathler, Mauricio Souza, Cid, André Maciel, Oliveira, Felipe, Freitas, Gustavo, Pessin, Gustavo and Azpúrua, Héctor
      In Simpósio Brasileiro de Automação Inteligente (SBAI) 2021