The Title Of Your Presentation
Resilience and Adaptation
in Robotics
The T-Resilience Algorithm
S. Koos, A. Cully and J.-B. Mouret
ISIR - UPMC/CNRS
«[We should not] wonder if some mishap may happen, but rather ask
what one will do about it when it occurs»
- Corbato, 2007
Photo M. Romero, 2011
How to recover from unanticiped failures?
Classic fault tolerance
Visinsky, 1994; Koren and Krishna, 2007
Unforseen situation?
Learning for resilience
- "Classic" reinforcement learning
- Direct policy search / BBO
- local optimization
- fast (20 minutes to one hour)
- Evolutionary algorithms
- more open search
- slower(several hours)
➔ Most of the time is spent in evaluating solutions
Togelius, 2009; Kohl and Stones, 2004; Peters and Schaal 2008; Hornby et al., 2005
Inspiration: Self-modeling
Bongard et al., 2006
T-Resilience: concept
- Optimize/learn in the self-model
- Avoid behaviors for which the self-model and the reality do not match
Transferability function
- A transferability function is a function that maps
descriptors of solutions to a transferability score that
represents how well the simulation matches the reality.
- Learned by supervised learning (SVM)
Koos, Mouret and Doncieux, 2012
T-Resilience algorithm
Experiments
- Performance: RBG-D SLAM algorithm (Kinect-like).
- Controller: 24 parameters (sinusoids)
Endres et al. 2011
Conclusion
- Current running time: 19 minutes, but can be reduced
- faster with faster machines
- half of the running time: rgb-D SLAM algorithm
- Tested on six different damages
- 1 to 3 times better than the other approaches
- A general approach for fast learning in robotics?
- Does something similar exist in biology?
Questions
mouret@isir.upmc.fr