Learning Disturbances in Autonomous Excavation. Maeda, G. J., Rye, D. C., & Singh, S. P. N. In Proceedings of the International Conference on Intelligent Robots and Systems (IROS), 2012. Submitted
doi  abstract   bibtex   
Disturbances that arise in material removal by repeated attempts to track the same path have the particular characteristics of non-repetitive magnitudes, but nearly-repetitive or gradual gradient transitions. This paper proposes and validates Iterative Learning Control (ILC) with a PD-type learning function for this class of disturbance as a predictive controller for autonomous excavation. However, parameters of the PD learning function may require different tunings for different excavation conditions, and convergence can be slow when compared to changes in excavation dynamics. In order to improve convergence, a plant inversion learning function is reinterpreted as a disturbance observer in the iteration domain, effectively rendering a disturbance learning controller (DLC). A hydraulic mini-excavator was used to evaluate experimentally the performance of the conventional ILC and the DLC against a robust controller. ILC achieved a desired cut profile with non-monotonic transients and DLC converged faster by learning disturbances directly from command discrepancies.
@INPROCEEDINGS{gjm.iros2012.draft,
  author = {Guilherme J. Maeda and David C. Rye and Surya P. N. Singh},
  title = {Learning Disturbances in Autonomous Excavation},
  booktitle = {Proceedings of the International Conference on Intelligent Robots
	and Systems ({IROS})},
  year = {2012},
  note = {Submitted},
  abstract = {Disturbances that arise in material removal by repeated attempts to
	track the same path have the particular characteristics of non-repetitive
	magnitudes, but nearly-repetitive or gradual gradient transitions.
	This paper proposes and validates Iterative Learning Control (ILC)
	with a PD-type learning function for this class of disturbance as
	a predictive controller for autonomous excavation. However, parameters
	of the PD learning function may require different tunings for different
	excavation conditions, and convergence can be slow when compared
	to changes in excavation dynamics. In order to improve convergence,
	a plant inversion learning function is reinterpreted as a disturbance
	observer in the iteration domain, effectively rendering a disturbance
	learning controller (DLC). A hydraulic mini-excavator was used to
	evaluate experimentally the performance of the conventional ILC and
	the DLC against a robust controller. ILC achieved a desired cut profile
	with non-monotonic transients and DLC converged faster by learning
	disturbances directly from command discrepancies.},
  doi = {10.1109/IROS.2012.6385566},
  pdf = {gjm.iros2012.draft.pdf},
  review = {http://db.acfr.usyd.edu.au/download.php/Maeda2012_IROS_LearningDisturbances.pdf?id=2571}
}

Downloads: 0