Navigating by touch: haptic Monte Carlo localization via geometric sensing and terrain classification

Abstract

Legged robot navigation in extreme environments can hinder the use of cameras and lidar due to darkness, air obfuscation or sensor damage, whereas proprioceptive sensing will continue to work reliably. In this paper, we propose a purely proprioceptive localization algorithm which fuses information from both geometry and terrain type to localize a legged robot within a prior map. First, a terrain classifier computes the probability that a foot has stepped on a particular terrain class from sensed foot forces. Then, a Monte Carlo-based estimator fuses this terrain probability with the geometric information of the foot contact points. Results demonstrate this approach operating online and onboard an ANYmal B300 quadruped robot traversing several terrain courses with different geometries and terrain types over more than 1.2 km. The method keeps pose estimation error below 20 cm using a prior map with trained network and using sensing only from the feet, leg joints and IMU.

Unsupervised Learning of Terrain Representations for Haptic Monte Carlo Localization

“…In [16], we presented a geometric only method based on Sequential Monte Carlo (SMC) Localization that estimated the 6-Degrees of Freedom (DoF) pose trajectory of the robot. Building on this work in [4], we added proprioceptive terrain classification to improve localization on flat grounds without distinct geometric features. However, this method required explicit classification of different terrains which may not always be practical in real scenarios.…”

“…3. This combination of blocks, inspired by the supervised terrain classification network from [4], extracts high-level feature representation of the time-series signals and can process variable-length data. The IAE model uses two residual layers with 1D convolutions in the CNN module and two bidirectional layers with two GRU cells in the recurrent component of an encoder.…”

“…The core localization algorithm used in this work is the same as [4]. We provide some explanation of the method here, however, details are omitted due to space constraints.…”

“…is the Euclidean distance between the foot's position (d i xf , d i yf ) and the matched, closest step in the sparse latent map (s i xf , s i yf ). We reject matches if the Euclidean distance exceeds the threshold of d t = 25 cm, which was set to match the similar Euclidean threshold for terrain classification in [4]. The sigma values of σ u = 0.4 determines the standard deviation of the F/T signal while sigma value of σ d = 0.4 captures the decreased credibility of the sparse map measurement when the foot was placed further away from the known location in the map.…”

“…It is possible to determine the class of terrain from tactile measurements [1], [2], [3], which in turn can be used to localize a legged robot in a prior map [4]. However, existing approaches are limited as they require a dense and terrainannotated map of the environment which often necessitates the use of external, ground truth sensors.…”