ARC-RL: Reinforcement Learning Environments for Game-Style Robot Morphologies

ARC-RL introduces a collection of four continuous-control environments based on MuJoCo, showcasing robotic designs inspired by the bestiary of the ARC Raiders game. The lineup features the 18-DoF tall hexapod named Queen, the 12-DoF armored hexapod called Bastion, the compact 18-DoF hexapod Tick, and the 12-DoF quadruped known as Leaper. Each robot utilizes a consistent observation framework, action protocol, simulation timing, and a singular closed-form reward function that varies slightly in weights and parameters across morphologies. This reward combines a velocity-tracking component with a survival bonus. This initiative seeks to bridge the divide between legged locomotion studies, which typically utilize real-world robotic designs, and game NPCs that often lack real-world equivalents. The environments serve as a testing ground for reinforcement learning, highlighting stylistic constraints not present in traditional sim-to-real robotics.