Uploaded May 2018 | Updated September 2026, 3 weeks ago
ICRA 2018 Spotlight Video
Interactive Session Wed PM Pod R.8
Authors: Wang, Chaoqun; Li, Teng; de Silva, Clarence; Meng, Max Q.-H.
Title: Efficient Mobile Robot Exploration with Gaussian Markov Random Fields in 3D Environments
Abstract:
In this paper, we study the problem of autonomous exploration in unknown environments. We use mutual information (MI) to evaluate the information the robot would get at a certain location. In order to get the best sensing location, we first proposed a sampling method that can get random sensing patches in free space. The sensing patch is grown to informative locations to collect information with true values. Then we use Gaussian Markov Random Field (GMRF) to model the distribution of MIs in the environment. Comparing with the traditional method that employ Gaussian Process (GP) to model the environment, GMRF is more efficient. MI of every candidate location can be estimated using the training sample patches we get and the GMRF model. We utilize an efficient computation algorithm to estimate the coefficients of GMRF so as to speed up the computation. Besides the MI of the candidates regions, the path cost is also considered in this work. We proposed a utility function that can balance the path cost and the information the robot would collect. We test our algorithm in both simulated and real experiment.The experiment results demonstrate that our proposed method can explore the environment efficiently with relatively shorter path length.
ICRA 2018 Spotlight Video
Interactive Session Wed PM Pod R.8
Authors: Wang, Chaoqun; Li, Teng; de Silva, Clarence; Meng, Max Q.-H.
Title: Efficient Mobile Robot Exploration with Gaussian Markov Random Fields in 3D Environments
Abstract:
In this paper, we study the problem of autonomous exploration in unknown environments. We use mutual information (MI) to evaluate the information the robot would get at a certain location. In order to get the best sensing location, we first proposed a sampling method that can get random sensing patches in free space. The sensing patch is grown to informative locations to collect information with true values. Then we use Gaussian Markov Random Field (GMRF) to model the distribution of MIs in the environment. Comparing with the traditional method that employ Gaussian Process (GP) to model the environment, GMRF is more efficient. MI of every candidate location can be estimated using the training sample patches we get and the GMRF model. We utilize an efficient computation algorithm to estimate the coefficients of GMRF so as to speed up the computation. Besides the MI of the candidates regions, the path cost is also considered in this work. We proposed a utility function that can balance the path cost and the information the robot would collect. We test our algorithm in both simulated and real experiment.The experiment results demonstrate that our proposed method can explore the environment efficiently with relatively shorter path length.
