Learning Optimal Decision Making for an Industrial Truck Unloading Robot using Minimal Simulator Runs
This work addresses data efficiency in robotic simulation for industrial automation, though it is incremental as it applies existing PAC frameworks to a specific domain.
The paper tackles the problem of an industrial robot learning to maximize boxes unloaded per action in truck unloading, using a high-fidelity but time-consuming simulator, and shows that their method achieves a significant reduction in simulator runs compared to naive approaches.
Consider a truck filled with boxes of varying size and unknown mass and an industrial robot with end-effectors that can unload multiple boxes from any reachable location. In this work, we investigate how would the robot with the help of a simulator, learn to maximize the number of boxes unloaded by each action. Most high-fidelity robotic simulators like ours are time-consuming. Therefore, we investigate the above learning problem with a focus on minimizing the number of simulation runs required. The optimal decision-making problem under this setting can be formulated as a multi-class classification problem. However, to obtain the outcome of any action requires us to run the time-consuming simulator, thereby restricting the amount of training data that can be collected. Thus, we need a data-efficient approach to learn the classifier and generalize it with a minimal amount of data. A high-fidelity physics-based simulator is common in general for complex manipulation tasks involving multi-body interactions. To this end, we train an optimal decision tree as the classifier, and for each branch of the decision tree, we reason about the confidence in the decision using a Probably Approximately Correct (PAC) framework to determine whether more simulator data will help reach a certain confidence level. This provides us with a mechanism to evaluate when simulation can be avoided for certain decisions, and when simulation will improve the decision making. For the truck unloading problem, our experiments show that a significant reduction in simulator runs can be achieved using the proposed method as compared to naively running the simulator to collect data to train equally performing decision trees.