1 / 19

Evolutionary Robotics

Evolutionary Robotics. Transferability. Question: Could the transferability of a controller be added to the fitness function? How to define transferability? Solution: .

tdavila
Download Presentation

Evolutionary Robotics

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript

  1. Evolutionary Robotics Transferability Question: Could the transferability of a controller be added to the fitness function? How to define transferability? Solution: Koos, S. and Mouret, J.-B. and Doncieux, S. (2010). Crossing the Reality Gap in Evolutionary Robotics by Promoting Transferable Controllers. GECCO'10: Proceedings of the 12th annual conference on Genetic and evolutionary computation.

  2. Evolutionary Robotics Multi-objective optimization (MOO) Objective 2: Behavior Objective 1: Transferability dominated : there exists another solution, t, for which all objectives ot >= all objectives os solution s = non-dominated : otherwise Koos, S. and Mouret, J.-B. and Doncieux, S. (2010). Crossing the Reality Gap in Evolutionary Robotics by Promoting Transferable Controllers. GECCO'10: Proceedings of the 12th annual conference on Genetic and evolutionary computation.

  3. Evolutionary Robotics Multi-objective optimization (MOO) x = controller at generation i = non-dominated controller at generation i+1 = dominated controller at generation i+1 Koos, S. and Mouret, J.-B. and Doncieux, S. (2010). Crossing the Reality Gap in Evolutionary Robotics by Promoting Transferable Controllers. GECCO'10: Proceedings of the 12th annual conference on Genetic and evolutionary computation.

  4. Evolutionary Robotics Multi-objective optimization (MOO) Q: Why low distance  low disparity? x = controller at generation i = non-dominated controller at generation i+1 = dominated controller at generation i+1

  5. Evolutionary Robotics Transferability Question: Could the transferability of a controller be added to the fitness function? How to define transferability? Solution: Use two fitness objectives: One that rewards for the desired behavior, another that punishes for differences between simulated and real behavior (transferability) Problem 1: …but to compute transferability for each evolved controller, have to evaluate each on the real robot! Koos, S. and Mouret, J.-B. and Doncieux, S. (2010). Crossing the Reality Gap in Evolutionary Robotics by Promoting Transferable Controllers. GECCO'10: Proceedings of the 12th annual conference on Genetic and evolutionary computation.

  6. Evolutionary Robotics Multi-objective optimization (MOO) Objective 2: Behavior Objective 1: Transferability Koos, S. and Mouret, J.-B. and Doncieux, S. (2010). Crossing the Reality Gap in Evolutionary Robotics by Promoting Transferable Controllers. GECCO'10: Proceedings of the 12th annual conference on Genetic and evolutionary computation.

  7. Evolutionary Robotics Multi-objective optimization (MOO) Objective 2: Behavior Objective 1: Transferability Koos, S. and Mouret, J.-B. and Doncieux, S. (2010). Crossing the Reality Gap in Evolutionary Robotics by Promoting Transferable Controllers. GECCO'10: Proceedings of the 12th annual conference on Genetic and evolutionary computation.

  8. Evolutionary Robotics Multi-objective optimization (MOO) Objective 2: Behavior Objective 1: Transferability Koos, S. and Mouret, J.-B. and Doncieux, S. (2010). Crossing the Reality Gap in Evolutionary Robotics by Promoting Transferable Controllers. GECCO'10: Proceedings of the 12th annual conference on Genetic and evolutionary computation.

  9. Evolutionary Robotics Multi-objective optimization (MOO) “Similar” behavior ? Objective 2: Behavior “Different” behavior ? Objective 1: Transferability Koos, S. and Mouret, J.-B. and Doncieux, S. (2010). Crossing the Reality Gap in Evolutionary Robotics by Promoting Transferable Controllers. GECCO'10: Proceedings of the 12th annual conference on Genetic and evolutionary computation.

  10. Evolutionary Robotics Transferability Question: Could the transferability of a controller be added to the fitness function? How to define transferability? Solution: Use two fitness objectives: One that rewards for the desired behavior, another that punishes for differences between simulated and real behavior (transferability) Problem 1: …but to compute transferability for each evolved controller, have to evaluate each on the real robot! Solution 1: Estimate transferability of a controller based on behavioral similarity to a tested controller. Koos, S. and Mouret, J.-B. and Doncieux, S. (2010). Crossing the Reality Gap in Evolutionary Robotics by Promoting Transferable Controllers. GECCO'10: Proceedings of the 12th annual conference on Genetic and evolutionary computation.

  11. Evolutionary Robotics Transferability Question: Could the transferability of a controller be added to the fitness function? How to define transferability? Solution: Use two fitness objectives: One that rewards for the desired behavior, another that punishes for differences between simulated and real behavior (transferability) Problem 1: …but to compute transferability for each evolved controller, have to evaluate each on the real robot! Problem 2: …how to measure the similarity between behaviors? Koos, S. and Mouret, J.-B. and Doncieux, S. (2010). Crossing the Reality Gap in Evolutionary Robotics by Promoting Transferable Controllers. GECCO'10: Proceedings of the 12th annual conference on Genetic and evolutionary computation.

  12. Evolutionary Robotics Transferability Problem 2: …how to measure the similarity between behaviors? Solution 2: Define a distance measure between any two behaviors. Define behavior b as a vector: b = (f1, f2, f3) f1 = distance traveled by robot f2 = mean height of the robot during travel f3 = final orientation of the robot ci = controller i; bi = behavior produced by controller i. Question: How to define the transferability of controller i? Koos, S. and Mouret, J.-B. and Doncieux, S. (2010). Crossing the Reality Gap in Evolutionary Robotics by Promoting Transferable Controllers. GECCO'10: Proceedings of the 12th annual conference on Genetic and evolutionary computation.

  13. Evolutionary Robotics Transferability Problem 2: …how to measure the similarity between behaviors? Solution 2: Define a distance measure between any two behaviors. Define behavior b as a vector: b = (f1, f2, f3) f1 = distance traveled by robot f2 = mean height of the robot during travel f3 = final orientation of the robot ci = controller i; bi = behavior produced by controller i. Question: How to define the transferability of controller i? Solution: c(s)i = controller i used in (s)imulation c(r)i = controller i used in (r)eality Koos, S. and Mouret, J.-B. and Doncieux, S. (2010). Crossing the Reality Gap in Evolutionary Robotics by Promoting Transferable Controllers. GECCO'10: Proceedings of the 12th annual conference on Genetic and evolutionary computation.

  14. Evolutionary Robotics Transferability Problem 1: …but to compute transferability for each evolved controller, have to evaluate each on the real robot! Solution 1: Estimate transferability of a controller based on behavioral similarity to a tested controller. Estimated transferability of controller c = CT = set of already-transferred controllers Koos, S. and Mouret, J.-B. and Doncieux, S. (2010). Crossing the Reality Gap in Evolutionary Robotics by Promoting Transferable Controllers. GECCO'10: Proceedings of the 12th annual conference on Genetic and evolutionary computation.

  15. Evolutionary Robotics Transferability Problem 1: …but to compute transferability for each evolved controller, have to evaluate each on the real robot! Solution 1: Estimate transferability of a controller based on behavioral similarity to a tested controller. Problem 2: …how to measure the similarity between behaviors? Solution 2: Define a distance measure between any two behaviors. Problem 3: Which controller to try in reality next? Koos, S. and Mouret, J.-B. and Doncieux, S. (2010). Crossing the Reality Gap in Evolutionary Robotics by Promoting Transferable Controllers. GECCO'10: Proceedings of the 12th annual conference on Genetic and evolutionary computation.

  16. Evolutionary Robotics Transferability Problem 1: …but to compute transferability for each evolved controller, have to evaluate each on the real robot! Solution 1: Estimate transferability of a controller based on behavioral similarity to a tested controller. Problem 2: …how to measure the similarity between behaviors? Solution 2: Define a distance measure between any two behaviors. Problem 3: Which controller to try in reality next? Solution 3: The one most “different” from those in the population. Koos, S. and Mouret, J.-B. and Doncieux, S. (2010). Crossing the Reality Gap in Evolutionary Robotics by Promoting Transferable Controllers. GECCO'10: Proceedings of the 12th annual conference on Genetic and evolutionary computation.

  17. Evolutionary Robotics Transferability Problem 1: …but to compute transferability for each evolved controller, have to evaluate each on the real robot! Solution 1: Estimate transferability of a controller based on behavioral similarity to a tested controller. Problem 2: …how to measure the similarity between behaviors? Solution 2: Define a distance measure between any two behaviors. Problem 3: Which controller to try in reality next? Solution 3: The one most “different” from those in the population. CT = already-transferred controllers Koos, S. and Mouret, J.-B. and Doncieux, S. (2010). Crossing the Reality Gap in Evolutionary Robotics by Promoting Transferable Controllers. GECCO'10: Proceedings of the 12th annual conference on Genetic and evolutionary computation.

More Related