Complex Terrain Navigation via Model Error Prediction

Abstract

Robot navigation traditionally relies on building an explicit map that is used to plan collisionfree trajectories to a desired target. In deformable, complex terrain, using geometric-based approaches can fail to find a path due to mischaracterizing deformable objects as rigid and impassable. Instead, we learn to predict an estimate of traversability of terrain regions and to prefer regions that are easier to navigate (e.g., short grass over small shrubs). Rather than predicting collisions, we instead regress on realized error compared to a canonical dynamics model. We train with an on-policy approach, resulting in successful navigation policies using as little as 50 minutes of training data split across simulation and real world. Our learningbased navigation system is a sample efficient shortterm planner that we demonstrate on a Clearpath Husky navigating through a variety of terrain including grassland and forest.

TerraPN: Unstructured Terrain Navigation using Online Self-Supervised Learning

“…These works are typically trained in a supervised manner using human-annotated image datasets. There have also been unsupervised learning-based works that automatically label terrains by characterizing the robot-terrain interaction using other sensor data such as forces/torques [40], vibrations and differences in odometry [14], [15], acoustic data [41], vertical acceleration experienced [42], and stereo depth [43], [44], etc.…”

“…For instance, works in indoor navigation have focused on socially-compliant navigation behaviors in avoiding and overtaking dynamic pedestrians [5], [6], [7], [8], groups of people [9], keeping to one side in corridors which helps avoid oncoming humans [10], [11], etc. In outdoor navigation, works have predominantly focused on estimating terrain traversability using semantic segmentation [12], [3], [13], self-supervised learning [14], [15], etc, and detecting complex outdoor obstacles [16].…”

TerraPN: Unstructured Terrain Navigation using Online Self-Supervised Learning

“…Few methods have performed self-supervised regression [6], where for each pixel in an image, the corresponding force-torque measurements are predicted post training. [7] presents a method for labeling images with the difference between a robot's actual trajectory and predicted trajectory based on its dynamics model. Ordonez et al [25] model the interaction between wheeled/tracked robots with pliable outdoor vegetation by mapping RGB data with the resistive forces experienced by the robot.…”

“…Self-supervised regression [6], [7] to predict navigability costs using input and label data collected in the real world overcomes the aforementioned limitations. That is, an image (input) can be associated with data vectors collected through Fig.…”