Multi‑source millimeter‑wave (mmWave) observations can furnish human representations reusable across sensing tasks, yet sparse detections and multipath effects make these observations unstable, limiting existing point‑cloud pretraining approaches.\
We introduce PT³, a prototype‑guided self‑supervised pretraining framework. It encodes local states within spatial regions as sparse compositions of shared prototypes and learns their temporal evolution via autoregressive prediction. To further mitigate radar perturbations, an observation‑consistency loss enforces representation stability across different viewpoints and noise conditions.\
After pretraining on unlabeled sequences from diverse sources, the frozen encoder is directly applied to downstream tasks—action classification, dense point generation, joint pose estimation, and future motion prediction—none of which were seen during pretraining. Across these benchmarks, PT³ outperforms prior pretraining baselines and often matches or exceeds task‑specific models.\
Review: PT³’s prototype sparsification and temporal autoregression effectively strengthen mmWave point‑cloud representations, delivering a robust and versatile feature encoder for multiple downstream applications.