Learning
Reusable Task Components using
Hierarchical Activity Grammars with Uncertainties
Kyuhwa Lee, Tae-Kyun Kim and Yiannis Demiris
Hierarchical Activity Grammars with Uncertainties
Kyuhwa Lee, Tae-Kyun Kim and Yiannis Demiris
Overview
 |
The learning of reusable action components of a task from human demonstrations is useful for various robotic applications, such as imitation learning. In this project, we present a novel activity grammar learning method capable of learning reusable task components from a reasonably small number of samples under noisy conditions. Our linguistic approach aims to extract the hierarchical structures of actions which can be recursively applied to help recognize unforeseen, more complicated tasks that share the same task components. |
Approach
 |
Key Features: 1) Search for frequently occurring action symbols that are subset of input samples to effectively discover the hierarchy. 2) Explicitly take into account the uncertainty values associated with input symbols due to the ambiguity inherent in low-level detectors. |
Experiment: The Towers of Hanoi
Results
Based on the learned model,
contextually inconsistent actions (detection error) are corrected. For
example, if a disk was observed as being lifted without dropping first,
a droppnig action is added before lifting in the parsing stage. Pure
imitation simply follows what has been observed without considering
constraints. Please see the following video to see an example.
Video Demonstration
 |
Training: Learn an activity grammar from each human demonstrator solving the puzzle using only 2 and 3 disks, repeated 3 times each, using our stochastic grammar learning method. Testing: Imitate a human demonstrator solving the puzzle using 4 disks which requires a much longer sequence of actions. Observed actions are parsed using learned activity grammar and corrected if inconsistent. Imitation is successful only if the puzzle is solved after executing parsed symbols. |
Results
 |
Evaluate
our method in low (L: indoor) and high (H: sunlight) noise conditions: LL, LH: Train on L, Test on L and H, resp. HL,HH: Train on H, Test on L and H, resp. The left graph shows the average success rate from 75 evaluations on each scenario. Note that single execution error of over 40 actions is considered failure which makes this experiment non-trivial. |
Video Demonstration
Acknoweldgement
Authors thank to anonymous reviewers for their constructive feedback. Also, thanks to Harold Soh, Dimitri Ognibene, Miguel Sarabia, Sotirios Chatzis, and Raquel Espinoza for their helpful discussion. This research was partially supported by the EU FP7 project EFAA (FP7-ICT-270490).