DRIVE: Data-driven Robot Input Vector Exploration

TL;DR

This work proposes Data-driven Robot Input Vector Exploration (DRIVE), a protocol that enables characterizing uncrewed ground vehicles (UGVs) input limits and gathering empirical model training data, and proposes a novel learned slip approach outperforming similar acceleration learning approaches.

Abstract

An accurate motion model is a fundamental component of most autonomous navigation systems. While much work has been done on improving model formulation, no standard protocol exists for gathering empirical data required to train models. In this work, we address this issue by proposing Data-driven Robot Input Vector Exploration (DRIVE), a protocol that enables characterizing uncrewed ground vehicles (UGVs) input limits and gathering empirical model training data. We also propose a novel learned slip approach outperforming similar acceleration learning approaches. Our contributions are validated through an extensive experimental evaluation, cumulating over 7 km and 1.8 h of driving data over three distinct UGVs and four terrain types. We show that our protocol offers increased predictive performance over common human-driven data-gathering protocols. Furthermore, our protocol converges with 46 s of training data, almost four times less than the shortest human dataset gathering protocol. We show that the operational limit for our model is reached in extreme slip conditions encountered on surfaced ice. DRIVE is an efficient way of characterizing UGV motion in its operational conditions. Our code and dataset are both available online at this link: https://github.com/norlab-ulaval/DRIVE.

Figures (7)

• Figure 1: Vehicle and terrain characterization done through DRIVE. The manufacturer-defined Naive input-space region is drawn in gray. The vehicle's true input-space, characterized through internal measurements, is shown in orange. Typical human driving is shown in red. The resulting body velocities are represented in green for gravel and blue for snow.
• Figure 2: Top view drawing of a SSMR. In orange is the commanded body velocity $^{\mathcal{R}}\bm{f}$ and the resulting body velocity $^{\mathcal{R}}\bm{v}$ is shown in green. The input-space $\mathcal{J}$ is shown in orange and the body velocity space $\mathcal{B}$ is shown in green. The difference between commanded and resulting body velocity is represented as the slip velocity $^{\mathcal{R}}\bm{g}$ in purple. All represented velocities have an angular component $(\cdot)_\omega$. Robot parameters are the wheel radius $r$ and vehicle width $b$.
• Figure 3: Commanded, encoder-measured and modeled wheel velocities for both sides of a SSMR during two DRIVE training intervals. The powertrain model is described in \ref{['sec:powertrain']}. Each training step consists of one transient-state window (in light gray) and two steady-state windows (in dark gray). Commands and measurements on the x-axis are acquired at a rate of 20Hz.
• Figure 4: Three different commercial platforms that were used for the experimental work: a Superdroid HD2 (1), a Clearpath Robotics Husky (2), and a Clearpath Robotics Warthog mounted on wheels (3). The platforms weigh 80kg, 75kg and 470kg, respectively.
• Figure 5: Data-gathering protocol performance for the HD2 on snow experiment. The top subplot illustrates the three data-gathering methods compared in this work. In yellow, we have the linear-focused method. In teal, we have the angular-focused method. In blue-violet is our DRIVE approach. The crosses and regions on the bottom subplot show the medians and interquartile ranges for translational and angular prediction errors.