Moss Robotics

Perception Intern · Summer & Fall 2026

Perception and tools for sensor kits used on tree farms.

Moss website ↗

Experiments

Isometric point cloud of tree rows with 3D detection boxes.

Synthetic training data

Compared real-only, synthetic-only, and mixed training data for tree detection.
The models reached diminishing returns with much less real data than I expected.
Synthetic examples helped most when real data was very scarce; with more real data, their added benefit was small or absent.
The synthetic dataset was still useful in other model experiments.

Model Architecture

Experimented with transformer-based tree detectors using 3D scan data.
Compared input sizes and training configurations across farm types, looking at the trade-off between finding small and large trees.

3D data and normalization

Studied how different ways of representing 3D LiDAR scan data influenced tree detection.
Experimented with unifying normalization settings that varied by farm layout and production type.
Tested point dropout augmentations to make models more robust to changes in point density.

Finding tree rows

Built an algorithm to detect tree rows, including curved and unevenly spaced ones.
Removed the need to annotate a reference row before estimating its curvature.

Demo of row detection in the inventory interface.

Masking the vehicle in LiDAR

Developed automatic masking to remove the sensor vehicle from LiDAR scans before mapping.
Adapted the method to different vehicles.

LiDAR comparison: vehicle points highlighted in orange on the left, with those points removed on the right.

Tree height detection

Worked on a model to refine tree-height estimates from 3D detections.
Estimated the tree’s own vertical extent when a full detection box included a stake or neighboring canopy.
Pretrained with a large synthetic dataset, then fine-tuned on real data.

Before: the recorded tree point cloud with a full detection-box estimate of 3.77 meters in the detector's coordinate frame.

After: the identical point cloud and camera, with a height model estimate of 1.44 meters in the detector's coordinate frame. Points outside that span remain visible in gray.

3D labeling tool

Built a labeling platform that brought 2D and 3D annotations into one workflow.
Added model assistance, review queues, dataset versioning, and exports.

Mabel labeling workspace with a recorded tree point cloud and 3D tree annotation boxes.

Mission data management

Built a mission-data catalog for uploads and metadata search.
Organized recordings with tags and groups.

Moss Raw Assets catalog with mission filters, metadata conditions, and grouped field recordings.

Other work

Built a camera-focus tool with per-corner sharpness and tilt feedback.