Nothing here is pre-recorded and there is no server: the world, the sensors and every algorithm are written from scratch in TypeScript and run in this tab, rendered with three.js.
What’s running
01Localization
Monte Carlo Localization with 800 particles against a surveyed prior map, using a likelihood-field laser model, low variance resampling and random particle injection. Tree rows look identical to a laser, so a ±1.5 m GNSS fix is fused in to break the symmetry. Kidnap the robot to watch it recover.
02Occupancy grid mapping
A 10 cm log-odds grid updated with an inverse sensor model: Bresenham ray tracing marks free space along each of the 360 laser beams, endpoints become obstacles. Each occupied cell also keeps the mean of its hits for sub-cell accuracy.
03SLAM
Scan-to-map SLAM: drifting wheel odometry predicts the motion, ICP against the map built so far corrects it, and the corrected scan is fused into the map. Compare the red odometry track with the green SLAM track and the white ground truth.
04Point clouds & ICP
A simulated 16-channel LiDAR ray casts trunks, porous canopies, fruit and the ground. Scans are registered with the SLAM pose into a voxel-filtered global cloud with semantic labels. The 2D ICP is point-to-point with a shrinking gate, Huber weights and a closed-form solve.
05Autonomous exploration
Frontier-based exploration: frontiers between known free space and unknown space are clustered, scored by travel cost and size, and reached with Dijkstra/A* on an inflated costmap and a pure pursuit controller, with stuck detection and recovery.
063D fruit mapping
Two side-looking RGB-D cameras detect fruit that faces them and isn’t hidden by leaves, with noise growing with depth. Detections are put in the map frame with the SLAM pose and associated in a spatial hash; a fruit counts once it’s been seen three times. Precision and recall are measured against the ground truth.