Incremental Task Modification via Corrective Demonstrations

摘要: In realistic environments, fully specifying a task model such that robot can perform in all situations is impractical. this work, we present Incremental Task Modification via Corrective Demonstrations (ITMCD), novel algorithm allows to update learned by making use of corrective demonstrations from an end-user its environment. We propose three different types updates make structural changes finite state automaton (FSA) representation the first converting FSA into transition auto-regressive hidden Markov (STARHMM). The STARHMM's probabilistic properties are then used approximate Bayesian selection choose best update, if any. evaluate ITMCD Model Selection simulated block sorting domain and full on real-world pouring task. simulation results show our approach new models sufficiently incorporate while remaining as simple possible. performs well when modeled segments closely comply with original model.