Pick one of the other approaches for understanding commands (Chen & Mooney, Winograd, Matuszek et al., or Branavan et al.,) and describe how it could be integrated with SLAM. What are the strengths and weaknesses of these representations for understanding language, compared to the approach described in this week's paper?
To me the most natural approach would be to combine the Branavan et al. approach with a SLAM problem. The robot can be given written directions, or spoken and the robot just tries to follow these actions, while confirming that its actions are leading to environments congruent to the spoken directions, and getting step wise rewards. This can be easy on loop closures as well, as directions would label the loop closure and there would be a payoff to achieve this loop closure. There would be a need of a semantic graph in this case as well, as most actions as described in the written map would not be precise. Teh strength of this approach would be that it could allow the robot to autonomously generate the metric map of the area more precisely without a user controlling/ following it. The downsides can be that the written maps would have to be generated (unlike MS manuals) and might not be precise. About the results of such a model compared to the approach in Walter et al. I can't being to stipulate.
ReplyDelete