@inproceedings{Tang.2026.Xplore.GNN,
  abstract   = "Transformable-wheel robots occupy a middle ground between conventional wheeled and legged systems: they can roll efficiently on even terrain, yet reconfigure to better handle obstacles. In this work, we study reinforcement learning for controlling a transformable-wheel robot and investigate whether a graph-structured actor-critic policy provides an advantage relative to a flat multilayer perceptron (MLP). Our policy represents the robot as one body node and four corner nodes, which encodes the platform’s symmetry and relational structure. We find in our experiments that graph-based policies improve early training behavior by helping the agent avoid certain local failure modes. Our results show that consistent forward locomotion can be learned and that both graph-based and MLP policies can perform well under the present task formulation, yet the graph-based policy is able to escape obstacle-relative local basins faster. We also identify the deep connection between reward task design and graph feature design for similar morphology dependent navigation tasks.",
  author     = "Tang, Tom and Chulani, Neil and Clark, Anthony J.",
  location   = "Vienna, Austria",
  publisher  = "MIT Press",
  booktitle  = "Cross-Disciplinary aspects of Exploration in Robotics, Reinforcement Learning and Search",
  date       = "2026-06-01",
  doi        = "",
  eventtitle = "Xplore at {ICRA} 2026",
  title      = "Graph-Structured Reinforcement Learning for Controlling a Transformable-Wheel Robot",
}