Robotic applications increasingly deal with more unstructured environments. Robots that can perceive and deal with uncertainty are much more robust in these scenarios.
Uncertainty arises from:
- unpredictability of the physical world
- limitations in sensor perception
- robot actuations involve motors that have control noise
- models of the world inherently inaccurate
- many algorithms are approximate
A robot that carries a notion of its own uncertainty that acts accordingly is superior to one that does not.
- Robust, and scale better to complex, unstructured environments
- Weaker requirements on the accuracy of the models compared to classical planning algorithms
- Sound methodology for many flavours of robot learning
- Broadly applicable to many problems, including perception and action
- Relatively computationally inefficient
- Requires approximation (exact posteriors are computationally intractable)
Thrun, S., Burgard, W., & Fox, D., Probabilistic robotics (2005), : MIT press. ↩