Abstract:
It is not feasible to pre-program robots a priori for every possible task they may encounter in unstructured domains. Upon encountering a task that a robot can’t solve, one common strategy is to teach it new skills via demonstrations. However, demonstrating a task can often be more cumbersome than performing the task directly. This provides a strong motivation for minimizing the number of demonstrations and ideally only teaching a skill to the robot if it will be useful for future tasks.

We formalize this problem as solving an MDP where we consider three meta-actions for every task: (1) the robot attempts the task using an existing skill (2) it skips the task and asks a human to do it and (3) it requests a human to provide demonstrations for the task. On the one hand, solving this MDP directly is intractable. On the other hand, greedy approaches for this problem can be highly sub-optimal. Using insights from the set cover problem we propose an efficient planner that seeks to find a sequence of meta-actions that minimizes the total execution effort, of both the robot and the human, to complete all of the tasks. In addition, our planner is theoretically guaranteed to do no worse than a greedy strategy. To capture the potential benefits of additional demonstrations, we train in simulation a novel parameterized precondition model that predicts which remaining tasks could be performed if a demonstration were provided for the current task. We validate our approach experimentally in simulation and in real world on sets of block and peg insertion tasks under sensing uncertainty. The experiments show a significant reduction in execution cost compared to alternative approaches.

Committee:
Prof. Maxim Likhachev (Advisor)
Prof. Oliver Kroemer (Advisor)
Prof. Chris Atkeson
Leonid Keselman