In a typical path planning pipeline for a ground robot, we build a map (e.g., an occupancy grid) of the envi- ronment as the robot moves around. While navigating indoors, a ground robot’s knowledge about the environment may be limited due to occlusions. Therefore, the map will have many as-yet-unknown regions that may need to be avoided by a con- servative planner. Instead, if a robot is able to correctly predict what its surroundings and occluded regions look like, the robot may be more efficient in navigation. In this work, we focus on predicting occupancy within the reachable distance of the robot to enable faster navigation and present a self-supervised proximity occupancy map prediction method, named ProxMaP. We show that ProxMaP generalizes well across realistic and real domains, improves the robot navigation efficiency in simulation by 12.40% against the traditional navigation method.
Here we show results from ProxMaP and its variations on living room data obtained from AI2THOR simulator. Abbreviations Reg and Class refer to Regression and Classification tasks, respectively. ProxMaP is a classification model.
Below we show results from ProxMaP and its variations on living room data obtained from Habitat-Matterport3D (HM3D) simulator. Abbreviations Reg and Class refer to Regression and Classification tasks, respectively. ProxMaP is a classification model.
Comparison of ProxMaP-Net and its variants against non-predictive navigation, an equivalent robot setup with 3 cameras, and another self-supervised occpancy map prediction method by Wei et al. in AI2THOR living rooms. We use Success weighted by (bornalized inverse) Completion Time (SCT) as the metric.
Method | SCT |
---|---|
Baseline (non-predictive) | 0.589 |
Wei et al. | 0.568 |
Reg-UNet (MSE) | 0.629 |
Reg-GAN | 0.592 |
Class-GAN | 0.632 |
ProxMaP | 0.662 |
3-Cameras (non-predictive) | 0.648 |
Below we comapre these methods in a more challenging situation by modifying FloorPlan227 to add more possible paths and thus emphasize decision-making. We find that ProxMap performs well compared to other predictive approaches and performs close to the from the 3-Camera setup.
Method | SCT |
---|---|
Baseline (non-predictive) | 0.615 |
Wei et al. | 0.623 |
Reg-UNet (MSE) | 0.571 |
Reg-GAN | 0.573 |
Class-GAN | 0.600 |
ProxMaP | 0.668 |
3-Cameras (non-predictive) | 0.690 |
Here we show prediction from ProxMaP over real data. We use a TurtleBot2 robot with Hokuyo laser scanner mounted on it. The field of view of the scanner is ;imited to 45 degrees and the data from it used to create the occupancy map. These examples show that even when only a part of the obstacles, boxese here, is visible, ProxMaP can extended in accurately and also predict the free area correctly.
Occupancy Map Inpainting for Online Robot Navigation. Minghan Wei, Daewon Lee, Volkan Isler, Daniel Lee. ICRA 2021.