Adaptive Task and Motion Planning for Object-Invariant Stacking Operations

This paper proposes a method for adapting an object-invariant action model for stacking operations. The proposed approach results in the generalization of stacking operations for multiple objects, in contrast to handcrafted action model-based task and motion planning (TAMP). Action model adaptation for new object stacking is done by drawing inferences using intuitions and learning a reward from visual clues based on the heuristically defined reward function. With this reward, the action model adaptation problem is formulated as an n-armed bandit problem. Additionally, an efficient action selection strategy is proposed for the n-armed bandit problem, which leads to fast convergence. The agent repetitively interacts with the environment in real-time to adapt the action model for the new object. After adaptation, the agent is capable of stacking objects with varying poses. The proposed approach is tested in simulation using the URS manipulator with a two-finger gripper. The proposed adaptive action model-based TAMP outperforms the traditional handcrafted action model-based TAMP by a significant margin in terms of success rate.