Project Overview

Dry-land ecosystems make up 40 percent of the global land surface and support nearly one-sixth of the world’s population. Restorebot aims to develop robotics systems capable of restoring degraded rangelands. We aim to have a robot independently identify sites of interest for restoration with shared autonomy capable of ingesting human knowledge to assist in site selection. In addition to learning more about restoration in these environments, we are also targeting seeding native grass species and placing Connectivity Modifiers or ConMods to help increase germination.

Our Platform

We utilize a Clearpath Husky A200 (Restorebot) with the following sensors; Ouster OS1-64 LiDAR (Light Detection and Ranging), a LORD Microstrain 3DM-GX5-15 VRU, a Trimble BX922 GPS with a Trimble AG25 antenna with 5cm RTK/RTX corrections, Intel RealSense D435 cameras, and Gig-E cameras. In May 2022 we only used one forward-facing D435 and one downward-facing Intel RealSense D435 camera. In November 2022 we modified the camera setup to include three outward-facing Gig-E cameras in combination with two downward-facing RealSense D435 cameras. All onboard processing is handled by a 32-core AMD Ryzen 3990x and a GTX 1650. The initial result of data collection and immediate post-processing yields cm-level accurate GPS localization, lidar-inertial localization results (as well as the point clouds and inertial measurements themselves), a 15cm-voxel Octomap of the plots, point clouds, and stereo pairs of images both of the robot’s immediate vicinity (roughly 225 degrees for the November 2022 data, and roughly 90 degrees for the May 2022 data) as well as the area in between the robot’s drive train.

We had two deployments using “RestoreCart”, the previous iteration of Restorebot. For these deployments, we only had a wheeled cart outfitted with a Nvidia Jetson Xavier, the Ouster OS1-64 LiDAR, PhidgetSpatial Precision 3/3/3 IMU, a Ublox NEO-M8P RTK GPS rover/base station pair, and one forward-facing and one downward-facing Intel RealSense D435 camera. We employed Trimble SPS855 GPS Base Station paired with a Trimble Zephyr Model 3 antenna to get a good estimate of the initial GPS coordinate, as well as the corners of each of the plots which were marked with rebar. In order to localize this data in the GPS frame, we mounted 32 cm diameter Radar Reflectors onto the rebar on the corners of each of the experimental plots. Obviously, the GPS estimates are not as precise as they are with our robot-mounted Trimble system, however it is still possible to localize the robot in the GPS coordinate frame with a larger margin of error.

Experiments and Results

So far, Restorebot has been deployed three times (November 2021, May 2022, and November 2022) to the Restoration Field site at the Canyonlands Research Center, hosted by the Redd family’s Dugout Ranch. Our ongoing strategy has been to deploy once in late spring to observe grass seed germination and again in the fall to observe grass seed success and maturation. This gives us an idea as to what factors are not only relevant to plant germination, but also their maturation and ultimate survival.

See our plots here

Restorebot has evolved over the course of our deployments as we have experimented with sensor configurations and improved our platform. In November 2021, we have sparse RTK GPS data, IMU data, point clouds, forward-facing and downward facing images. For May and November 2022 we have extremely high fidelity GPS data, IMU data, point clouds, forward-facing and downward facing images. These data help us to localize the robot as well as objects in its environment in the GPS coordinate frame, in some cases down to the centimeter level, allowing us to better associate plant outcomes over multiple seasons. By localizing the robot, we can attribute features in the environment as well as individual plants to a very precise GPS location, allowing us to make conclusions about plant outcomes.

Currently, we are working on terrain identification that enables the robot to avoid obstacles such as rebar and ConMods. Additionally, we aim to have the robot avoid shrubs that are already growing in the environment despite being considered traversable by most conventional terrain identification systems. We are working on utilizing Large Language Models (such as ChatGPT) to develop convenient interfaces for ecologists to interact with RestoreBot. This includes developing pipelines that enable the robot to determine planting patterns from Natural language and its onboard sensors.

Upcoming Deployments

We are diligently planning for our November 2023 deployment in which we will focus on making autonomous intervention for the first time. In addition to refining our seeding mechanism, we are also planning to deploy a U5 robotic arm for installing custom ConMods. The ConMods used by ecologists are currently made of chicken wire, but to make them more friendly for robotic deployment, we have attempted to make similar structures with rigid wooden frames.