Foundation models endow autonomous systems with richer visual understanding; a notable example is Human Mesh Recovery (HMR), which estimates a target's 3D pose and shape for tactical missions. However, the large size and power demand of such models hinder real‑time execution on edge platforms, especially for active perception where a robot must plan its next‑best view (NBV) onboard and cannot rely on off‑board computation under contested communications.
We introduce LEAP‑NBV, a lightweight active‑perception framework that runs foundation‑model‑driven NBV planning directly on an edge device. First, we distill a family of large HMR teachers into compact 32M‑parameter student models using an offline mesh objective. Then we quantize the vision encoder to FP16 and characterize its on‑device accuracy and latency.
Within an occlusion‑aware perception loop, all compression configurations are evaluated on the same held‑out benchmark, and the full pipeline is deployed on an NVIDIA Jetson Xavier NX. Distillation recovers 6‑7 mm of Procrustes‑aligned mean per‑vertex position error (PA‑MPVPE) over the undistilled student on the test set. Selecting the edge‑optimal compression model brings the HMR engine to ~12 ms inference time, and the closed‑loop runs at 3.6 FPS with 2.6 J per frame, achieving a 2.0× speedup and 3.0× lower energy consumption compared to the uncompressed model while nearly preserving downstream task quality.
Review: LEAP‑NBV demonstrates that careful distillation and quantization can dramatically reduce computation and energy costs without sacrificing accuracy, offering a practical solution for real‑time active perception on resource‑constrained edge robots.