Data-Efficient Visual Perception Under Edge-Computing Constraints: Application to Autonomous Localization and Humanoid Robot Soccer

For small robots, annotation can outlast computation as a bottleneck: hardware is bought once, labels at every deployment. We study label efficiency for landmark detection on a humanoid soccer robot. Against soft-argmax and pooled coordinate-regression baselines on 10–100% of 1193 labeled images under one protocol and two held-out sessions, the heatmap configuration is the more label-efficient: with hyperparameters fixed on the full pool, 298 training images (509 annotated) give 0.436 PCK@10 px, which neither baseline reaches with all 1193. The heatmap and soft-argmax arms differ in target, loss and readout, so the interval compares configurations, not representations; at equal held-out false-detection rates the heatmap stays ahead at both tolerances. A validation-to-held-out gap widens from 0.245 to 0.350 PCK@10 with more labels, similarly on both sessions. Int8 conversion, on the host CPU for two checkpoints, changes PCK@10 by at most 0.003. Two Edge TPU accelerators run the detectors in 5.18 and 19.71 ms, timed alone; reduced host bandwidth slows them by 2.8 and 8.2 for the landmark network despite near-complete parameter caching. A Monte Carlo filter is accurate to centimeters in simulation; one field demonstration agrees with hand-annotated positions only to decimeters, with an unresolved frame offset. Loop latency is unmeasured.