Australasian Conference on Robotics and Automation (ACRA 2026)
We study how reinforcement learning policies trained on individual skills can be combined and reused to solve longer tasks without expert demonstrations.
Navigation in MiniGrid FourRooms. A manually specified plan combines two learned skills, GoTo and CrossDoor. Across 10 previously unseen layouts, the combined skills achieve 55.5% mean success, compared with 2.0% for a flat DQN given the doorway coordinates.
Manipulation in Robosuite. The Reach → Grasp → Lift chain picks up and lifts a cube in 85.8% of 450 evaluation trials. Keeping these three policies frozen and training only an additional Place skill gives 71.6% success on cube stacking.