Hallucinating Robots: Inferring Obstacle Distances from Partial Laser Measurements

Abstract

Many mobile robots rely on 2D laser scanners for localization, mapping, and navigation. However, those sensors are unable to correctly provide distance to obstacles such as glass panels and tables whose actual occupancy is invisible at the height the sensor is measuring. In this work, instead of estimating the distance to obstacles from richer sensor readings such as 3D lasers or RGBD sensors, we present a method to estimate the distance directly from raw 2D laser data. To learn a mapping from raw 2D laser distances to obstacle distances we frame the problem as a learning task and train a neural network formed as an autoencoder. A novel configuration of network hyperparameters is proposed for the task at hand and is quantitatively validated on a test set. Finally, we qualitatively demonstrate in real time on a Care-O-bot 4 that the trained network can successfully infer obstacle distances from partial 2D laser readings.

Deep Network Uncertainty Maps for Indoor Navigation

“…In this section, we describe an approach for predicting uncertainty of distance estimates and how to integrate it when building occupancy maps. Following the formalism presented in [5], let us define the output of a generic N -point 2D laser positioned at a height h from floor level as a 1D vector…”

“…We recently shifted the focus from estimating obstacle occupancy from raw 2D laser data towards learning to infer robot-to-obstacle distances using neural networks [5]. The results demonstrated first of all that typical indoor environments include enough structure to learn the robotto-obstacle distance of objects such as tables and windows, and secondly that the learned robot-to-obstacle distance can improve local navigation safety.…”

“…The size of the layers' outputs are the opposite to the encoder's, i.e., 16 × 64, 32 × 32, 64 × 16, and 128 × 1. Corresponding convolutional and deconvolutional layers are connected through skip connections, as their ability to improve the performance of the network has been demonstrated in this domain [5]. The input of the network is scaled to the range [0, 1] to improve convergence speed, and weights are optimized using Adaptive Moment Estimation (Adam).…”

“…The occupancy probability P (c = 1|ŷ r ) of a cell c given an estimated robot-to-obstacle distanceŷ r along a laser ray r is defined as P (c = 1|ŷ r ) = ∞ 0 P (c = 1|y r )P (y r |ŷ r ) dy r , (5) where y r represents the true robot-to-obstacle distance along ray r. Here, we model the occupancy given the true robot-toobstacle distance P (c = 1|y r ) as 0 in front of the obstacle, 1 Calculating the integral in (5) requires a model P (y r |ŷ r ) of the true robot-to-obstacle distance given a noisy measurement reading. As stated in Section III, P (y r |ŷ r ) is modeled as a Laplace distribution.…”

“…We used the trained network H, whose training setup was detailed in [5]. Then, to train the uncertainty network U, we took the complete training and test set used in [5], available online 1 , and merged them.…”