Action selection for touch-based localisation trading off information gain and execution time

N. Tosi (1) , O. David (1) , H. Bruyninckx (2)

1 DIASI (CEA, LIST) - Département Intelligence Ambiante et Systèmes Interactifs
2 KU Leuven - Catholic University of Leuven = Katholieke Universiteit Leuven

In the context of touch-based object localisation, solving the problem of 'where to sense next' is a challenging task due to the curse of dimensionality related to belief-state reasoning. We present a constrained optimisation scheme that computes the next best action maximising the trade off between (i) the localisation information gain, (ii) the time required for its computation, and (iii) the motion-execution time. This allows the robot programmer to have a deterministic influence on the length of every sensing action. The proposed methodology is applied to localise a solid object in 3D using a Staubli RX90 robot equipped with a force-torque sensor coupled with a spherical end effector. A case-study comparison of the task executed with two different time constraints is presented.

Domaines: Sciences de l'ingénieur [physics]

N. Tosi, O. David, H. Bruyninckx. Action selection for touch-based localisation trading off information gain and execution time. 2014 IEEE International Conference on Robotics and Automation (ICRA), May 2014, Hong Kong, China. pp.2270-2275, ⟨10.1109/ICRA.2014.6907173⟩. ⟨cea-01834434⟩