Technology

Autonomy in challenging, unstructured environments

Mowing is a unique autonomy problem set that requires extremely high precision. Unlike self-driving cars, we intentionally try to come close to obstacles.

For example, knowing we need to mow a park with, say, 150 trees, grills, benches, flower beds, and people as obstacles means we have to be experts in localization, mapping, scene recognition, object detection, navigation, and safety.

Start with Safety

When heavy equipment is involved, there’s no taking chances. This requires a symphony of hardware and software communication to see obstacles well in advance, both objects small and large. Our Safety system is firstly LIDAR based, using the most reliable sensor on the market, combined with secondary systems for maximum safety and uptime.

The foundation of autonomy

Though autonomy itself is complex, we can simplify it into three main questions we need to answer for our robots, through advanced communication between hardware and software data:

Where am I?

What’s around me?

Where do I need to go?

Part 1:
Hardware

We start by retrofitting proven, high-performance equipment with our AI-sensor stack.

  1. GPS, Cameras, communication infrastructure, safety lights
  2. Weather-proof compute box
  3. Certified safety sensor
  4. Rear safety stop
  5. Primary LIDAR

Part 2:
Software

Localization

Advanced sensor fusion techniques pull together inputs from various sensors, providing the robot a precise and accurate view of where it is in the world, even in the presence of dynamic obstacles.

Mapping

Reconstructing a digital copy of the real world enables the robot to navigate and make optimal decisions in area coverage. Map maintenance is crucial, and enables our robots to keep their model coherent with the real world, accurately localize, and safely navigate evolving terrain.

Segmentation

Scene segmentation allows the robot to make good decisions operating around boundaries.
This creates a secondary system to assess where the edge of the mowable areas are (i.e. this is grass, this is not).

Persistent object segmentation allows us to mask out classified objects within the map that might move. For instance, a car parked along the road would be a bad object to localize from.

Technology-img-compare-slider-1 Technology-img-compare-slider-2

Navigation

Optimal terrain coverage – once the robot knows where it is and the obstacles around it, now it has to navigate within that environment, safely and efficiently, while leaving no grass un-cut in the process. We are able to maximize efficiency, by covering the area in the shortest time and at the highest quality.

Adjustable

Sites are constantly changing, from events happening on campus to flooded corners of the field. Our machine’s UX allows for that adaptability.

Plans around obstacles

In efficient ways, rather than discovering these obstacles throughout the day. This eliminates the surprises, and allows for efficient maneuverability around obstacles that aren’t going anywhere week-to-week.

Multiple striping patterns

To minimize rutting in the grass. We offer the operator a variety of different striping patterns to choose from.

Obstacle detection

Knowing what’s hiding in front of the machine is critical to safety and high performance. This requires the machine to understand the difference between, say, a leaf, a piece of paper, or a small turtle. One you can mow over, one you need to throw away, and one you need to protect. This requires effective communication between LIDAR and camera vision to make that distinction.