Automatic field boundary generation

An autonomous off-road vehicle (AOV) comprising:

• a spectral camera;
• a hardware processor;
• and
• a non-transitory computer-readable storage medium storing executable instructions that, when executed by the hardware processor, cause the AOV to perform steps comprising:
• capturing, via the spectral camera and while the AOV is being manually operated in a field, spectral image data representative of contents of the field, the spectral image data comprising visible spectrum image data and non-visible spectrum image data;
• training a machine-learned model using a training set of non-visible spectrum image data, the training set including:
• (i) non-visible spectrum images of historical fields labeled as including crops, (ii) non-visible spectrum images of the historical fields labeled as not including crops and further labeled as navigable, and (iii) non-visible spectrum images of the historical fields labeled as not including crops and further labeled as non-navigable;
• identifying, based on the captured non-visible spectrum image data and by applying the machine-learned model to the spectral image data, (i) portions of the field that include crops, (ii) portions of the field that do not include crops but are navigable, and (iii) portions of the field that do not include crops and are non-navigable, wherein the machine-learned model is configured to classify portions of the visible spectrum image data as including crops and not including crops but navigable or non-navigable based at least in part on the non-visible spectrum image data;
• setting a field boundary for the field representative of areas where the AOV can navigate based at least in part on the identified portions of the field;
• and
• operating the AOV in an automated mode, wherein the AOV, when operating in the automated mode, is configured to generate a navigation path through the field within the field boundary.

A method for autonomously navigating an off-road vehicle (AOV) comprising:

• capturing, via one or more sensors of an AOV being manually operated in a field, spectral image data representative of contents of the field, the spectral image data comprising visible spectrum image data and non-visible spectrum image data;
• capturing, via a location sensor mounted on the AOV, geolocation data corresponding to the spectral image data;
• transmitting the spectral image data and the corresponding geolocation data to a central server, the central server applying a machine-learned model to the spectral image data to identify, based on the captured non-visible spectrum image data, (i) portions of the field that include crops, (ii) portions of the field that do not include crops but are navigable, and (iii) portions of the field that do not include crops but are non-navigable, by classifying portions of the visible spectrum image data as including crops and not including crops but navigable or non-navigable based at least in part on the non-visible spectrum image data, wherein the machine-learned model is trained using a training set of non-visible spectrum image data, the training set including:
• (i) non-visible spectrum images of historical fields labeled as including crops, (ii) non-visible spectrum images of the historical fields labeled as not including crops and further labeled as navigable, and (iii) non-visible spectrum images of the historical fields labeled as not including crops and further labeled as non-navigable;
• receiving, from the central server, a field boundary representative of areas where the AOV can navigate based at least in part on the identified portions of the field;
• and
• operating the AOV in an autonomous operation mode by generating a navigation path through the field within the generated field boundary.