Hussain Habib.

Hive-Nav coordinates a fleet of DJI RoboMaster EP Core robots across a university wing. Each robot carries a Raspberry Pi 5 and runs its own Nav2 stack. A single GPU workstation orchestrates the fleet and holds the inference models. The mission runs in two phases.

Phase 1 is semantic room classification under dynamic allocation. The workstation dispatches a room task list, and each robot measures candidate path lengths locally with ComputePathToPose. Allocation therefore reflects true reachable distance rather than straight line distance, and no planning query ever crosses the domain bridge. On arrival a robot spins in place and captures frames. The workstation classifies the room with a tiered pipeline that starts at YOLO detection and escalates to CLIP and a vision language model when tier-one evidence is weak. The allocator confirms or denies each room until the floor is resolved.

Phase 2 is coverage search. Once every room is identified, the workstation computes viewpoint waypoints per room from an open-area flood fill and streams them to the robot, which drives NavigateThroughPoses and samples frames for object detection at each pose.

Platform

RoboMaster EP Core

Onboard compute

Pi 5 per robot

Navigation

Nav2 AMCL

Sensing

RPLiDAR A2M12

A small slice of the latent is supervised by a physiological anchor, such as leaf count, plant height, or leaf chlorophyll. The rest is a residual that a predictor carries forward in time. In a controlled comparison against SIGReg, both the anchor and the regulariser prevent collapse. The anchor is better on the traits it supervises, reaching an age error of 1.69 days against 1.91 to 2.46, and it needs no weight to tune.

Which anchors help in the field. Chlorophyll readings on greenhouse soil boxes cut prediction error relative to persistence from 0.44 to 0.25 in three of three seeds, and plant height on wheat plots raises held-out R squared from 0.570 to 0.665. A calendar anchor, days after planting, gives no reliable benefit across seasons. Six heterogeneous sources train under one encoder once each source's loss is normalised by its own persistence error. Pooling then matches solo accuracy and cuts seed-to-seed spread by about three times.

What it gives back. With no labels at all, clustering trajectory features recovers four wheat varieties across 20 field subplots with an adjusted Rand index of 0.89, where colour statistics, ImageNet features, and an untrained encoder reach 0.37 to 0.43. On the lab benchmark, four camera views lower held-out prediction error by 40 percent in 15 of 15 matched runs, and the gain peaks at four views.

Not everything worked, and I would rather say so here. The model does not learn a dynamics prior that transfers to a new crop without retraining. Conditioning on management actions did not help in any way I could measure. Several forecasting gains that looked good against persistence mostly disappeared once I compared against a simple clock baseline. Anyone building on this work should know where the model stops being useful.

Field anchor

0.44 to 0.25error vs persistence

Varieties, no labels

0.89adjusted Rand index

Pooled sources

63x tighter seed spread

Four views

40%lower prediction error