Solving Real Problems on a Real System
An independent implementation of imitation learning:
Using MMP with LEARCH optimization to learn Crusher’s navigation system.
David Silver, J. Andrew Bagnell
In some of our early experiments on MMP we applied the framework to navigational planning. We demonstrated that MMP can generalize disparate concepts given disparate training examples. When demonstrating road following behavior, MMP trains A* to produce paths in novel regions that follow roads. Alternatively, when demonstrating “stealthy” off-road behavior, MMP trains A* to find paths through covered, highly vegetative regions. Our original paper (Ratliff et al., 2006a) presents these results in detail along with results on urban navigational planning. See the MMP page for a more detailed discussion of this algorithm.
Later, at the National Robotics Engineering Center (NREC), researchers lead by David Silver and Dr. J. Andrew Bagnell applied these algorithms to a large-scale real-world overhead navigation problem (Silver et al. 2008). Using example paths collected via human teleoperation, they learn the baseline navigational (prior) cost map for outdoor autonomous navigation on the DARPA sponsored Crusher vehicle designed and constructed at NREC. MMP with LEARCH optimization successfully trained the state-of-the-art field D* planning algorithm (Ferguson & Stentz, 2005). This planner integrates interpolation technology with meticulous caching procedures to find smoother paths that can be easily executed on the robot, and to implement real-time re-planning capabilities that allow the robot to adapt to a changing environment. The algorithm plays a central role in the success of Crusher's autonomous navigation.
The video above shows the training process along with a real-world data-log of Crusher executing a navigational plan found using the learned cost map. The navigational cost map used in these videos was demonstrated in over 600km of sponsor monitored autonomous navigation and has undergone several more kilometers of extensive field testing (Silver et al. 2008).