Watching Grass Grow: Long-term Visual Navigation and Mission Planning for Autonomous Biodiversity Monitoring

TL;DR

This work tackles long-term autonomous biodiversity monitoring in a highly dynamic grassland by integrating vision-based localisation with an experience-graph representation, teach-and-repeat navigation, and topological mission planning, complemented by LiDAR-based safety and autonomous docking. Field validation occurred over six weeks at Wytham Woods, spanning 40 grassland plots within climate-manipulation experiments and yielding >14 km of autonomous traversal with multiple >1 h missions. Key contributions include a practical mapping workflow that stitches local sequences into a global experience graph and a higher-level supergraph for efficient mission scheduling, along with robust docking and charging for sustained operation. The results demonstrate the viability of scalable, repeatable autonomous data collection for outdoor plant biodiversity monitoring and climate-change experiments, with implications for agricultural robotics and environmental sensing in complex natural settings.

Abstract

We describe a challenging robotics deployment in a complex ecosystem to monitor a rich plant community. The study site is dominated by dynamic grassland vegetation and is thus visually ambiguous and liable to drastic appearance change over the course of a day and especially through the growing season. This dynamism and complexity in appearance seriously impact the stability of the robotics platform, as localisation is a foundational part of that control loop, and so routes must be carefully taught and retaught until autonomy is robust and repeatable. Our system is demonstrated over a 6-week period monitoring the response of grass species to experimental climate change manipulations. We also discuss the applicability of our pipeline to monitor biodiversity in other complex natural settings.

Figures (11)

• Figure 1: ClearPath Husky base platform (top) with Bumblebee2 stereo vision (facing forwards) camera used for mapping and localisation with Hokuyo lasers (front bumper) for obstacle avoidance and safety curtain. Pointing sideways are payload cameras, not used for navigation, including monocular, thermo and multi-spectral. Plant-growing experiment enclosures (bottom), where rainwater is fed into four quadrants at different rates to investigate grass species' response to climate change.
• Figure 2: Manual data collection at one of the 40.0 grassland plots. A researcher places a transparent box over the plants of interest, with interesting features to be drawn by hand.
• Figure 3: An aerial shot of the Wytham Woods experimental site, with overlays indicating growing areas of interest for autonomous monitoring of complex, dynamic grasslands. Wytham Woods are a 423.8-hectare biological Site of Special Scientific Interest north-west of Oxford in Oxfordshire. The topological map or "supergraph" (\ref{['sec:planning']}) for autonomous navigation is shown in black. For example, in \ref{['fig:hx_map']}, consider edges in the graph S-B16 and B16-B14 (all other node names are not shown, to avoid clutter). Each edge also represents sequences of images or visual "experiences" (\ref{['sec:dub4']}). Autonomous docking, charging, and data offload (\ref{['sec:docking']}) occurs at a tent (\ref{['sec:hardware']}) near node H.
• Figure 4: Vision-based mapping and localisation, based on linegar2015work. The robot frame of reference is shown as the red and green axes, just to the left of a path previously built by an odometry chain (white). Local ORB features rublee2011orb (orange and blue) are used to estimate precise pose with respect to the path. The framework supports multi-experience localisation (bottom right), with just $1$ experience in this example. Illumination-invariant image transformations mcmanus2014shady (right) are used for improve robustness to glare and shadows, etc.
• Figure 5: Autonomous docking LiDAR perception system. The white dots are LiDAR returns. The red lines are detected line segments. The orange ones are the location of the predefined model that was chosen to best match the detected line segments. The green lines are the desired locations of the docking station layout after docking is complete -- i.e. the control reference signal.