BRENNA D. ARGALL


EPFL Logo

TACTILE CORRECTIONS for POLICY REFINEMENT and REUSE

Demo
In my postdoctoral work, the use of corrective human feedback has been expanded to other robot domains and feedback techniques. In particular, online feedback is provided to a non-mobile, high degree of freedom humanoid through a tactile  interface located on the body of the robot. We posit that tactile feedback naturally extends the idea of teaching robots as humans teach other humans. Moreover, in the case of robot operation around humans, tactile detection can be crucial for safe robot operation, and these already-detected interactions may be exploited to also further knowledge transfer from human to robot [3].

CorrectCorrective human touch is used to indicated incremental position adjustments in Cartesian space. The adjustments are immediately adopted, thus producing new behavior examples for the policy. Tactile feedback is used both to refine the resulting policy, with the goal of more successful and robust behavior, as well as to assist in its reuse in the development of other policies. The goal of reuse is to further reduce the effort required in policy development, (i) by building on behavior knowledge already present in an existing policy and (ii) through the opportunity to reuse a single policy multiple times, in the development of multiple distinct policies. Our initial implementation has been validated with grasp positioning on a humanoid robot [2,6,7], as well as grasp adaptation in response to changing contact with an object [1]. Under this approach, data derives from a variety of teachers and teaching techniques, as well as for multiple target behaviors [video]. We also have begun consider the unique intersection of each to be a distinct data source, and to explicitly address different sources within our algorithm [8], and to learn a touch-to-correction mapping with a skin of tactile sensors [4].
Journal Publications

[1]   Eric L. Sauser, B. D. Argall, Giorgio Metta and A. G. Billard. Iterative Grasp Adaptation Learning with Tactile Corrections. Robotics and Autonomous Systems, 60(1), 55-71, 2012.   [pdf]  [video]

[2]   B. D. Argall, Eric L. Sauser and A. G. Billard. Tactile Guidance for Policy Adapatation. Foundations and Trends in Robotics, 1(2), 79-133, 2010.   [pdf]   [video]

[3]   B. D. Argall and A. G. Billard. A Survey of Tactile Human-Robot Interactions. Robotics and Autonomous Systems
. 58(10): 1159-1176, 2010.   [pdf]



Referreed Conference Publications


[4]  B. Argall and A. Billard. Learning from Demonstration and Correction via Multiple Modalities for a Humanoid Robot. In Proceedings of the International Conference SKILLS, Montpellier, France, December 2011.   [pdf] 

[5]   B. Argall, E. Sauser and A. Billard. Policy Adaptation through Tactile Feedback. In Proceedings of the Sixth Annual Conference on Human Robot Interactions (HRI '11), late-breaking report, Lausanne, Switzerland, March 2011.

[6]  B. Argall, E. Sauser and A. Billard. Tactile Feedback for Policy Refinement and Reuse. In Proceedings of the 9th IEEE International Conference on Development and Learning (ICDL '10), Ann Arbor, Michigan, August 2010.   [pdf]   [video]

[7]  
B. Argall, E. Sauser and A. Billard. Policy Adaptation through Tactile Correction. In Proceedings of the 36th Annual Convention of the Society for the Study of Artificial Intelligence and Simulation of Behaviour (AISB '10), Leicester, UK, March 2010.   [pdf]

[8] 
B. Argall, E. Sauser and A. Billard. Demonstration, Tactile Correction and Multiple Training Data Sources for Robot Motion Control. In NIPS 2009 Workshop on Learning from Multiple Sources with Application to Robotics, Whistler, British Columbia, December 2009.   [pdf]