LEAP-NBV: Lightweight Edge Active-Perception for Foundation-Model Next-Best-View Planning
Boxun Hu, Jiawei Ge, Axel Krieger, Peng Wang, Tinoosh Mohsenin
cs.AI
Quick verdict
Competent edge-deployment study: distilled HMR + FP16 yields real Jetson gains (3.6 FPS), but evaluation is synthetic-only, with no variance reporting.
Summary
LEAP-NBV targets on-board next-best-view (NBV) planning for mobile robots whose perception engine is human mesh recovery (HMR). The authors distill four large SMPLer-X/SMPLest-X teachers (327-687M parameters) into a single 32M ViT-S student using an offline, centroid-relative per-vertex mesh loss, then apply FP16 post-training quantization to every weighted layer except the SMPL-X tail. On a held-out synthetic benchmark (504 single-view crops, 441 closed-loop frames, 20-60% controlled occlusion), distillation recovers 6-7 mm PA-MPVPE over the undistilled 32M baseline (62.9 mm), with students near 56 mm versus teacher values of 47.3-48.5 mm. They also report a width-dependent FP16 overflow in the final normalization that collapses SX-h (47.3 to 128.8 mm) but not the D=384 student. The deployed FP16 student runs HMR in about 12 ms and the full pipeline on a Jetson Xavier NX at 3.6 FPS and 2.58 J per frame - 2.0x faster and 3.0x lower energy than the uncompressed model - while closed-loop task metrics (coverage 62% to 88%, area up ~2.5x) nearly match teachers. Limitations: evaluation is entirely synthetic, no variance statistics are reported, and distillation training details are incomplete.
Review complete
Assessment at a glance
4
risk flags identified
● high 0
● medium 2
● low 2
5 strengths · 4 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- Cross-scale offline mesh distillation compresses 327-687M HMR teachers into one 32M student, recovering 6-7 mm PA-MPVPE over the same-size undistilled baseline (62.9 -> ~56 mm).
- On-device FP16 post-training quantization yields ~12 ms HMR on the Jetson Xavier NX at roughly 1 mm accuracy cost versus the FP32 distilled student.
- A width-dependent FP16 overflow in the final normalization collapses large-embedding models (SX-h: 47.3 -> 128.8 mm) while the D=384 student is numerically safe.
- The deployed pipeline runs the closed loop at 3.6 FPS and 2.58 J/frame, 2.0x faster and 3.0x lower energy than the uncompressed l32 pipeline.
- Downstream closed-loop task quality (coverage 62% -> ~88%, ~2.5x reconstructed area) of the deployed student nearly matches the 327-687M teachers.
Strengths
Real end-to-end on-device deployment with measured energy
Sec. IV-D and Fig. 6 profile board power (VDD_IN) per pipeline stage on the Jetson Xavier NX, reporting 276 ms (3.6 FPS) and 2.58 J/frame versus 550 ms and 7.83 J for the uncompressed l32 pipeline; the HMR stage drops 257 ms -> 12 ms while other stages match, which is a clean ablation of where the savings come from.
Practical offline distillation design
Sec. III-B precomputes teacher meshes once per training crop, decoupling student training from teacher inference; the centroid-relative per-vertex loss (Eq. 4) removes camera translation and matches the Procrustes-aligned metric, a sensible objective choice for a label-free, multi-teacher recipe.
Useful and falsifiable FP16 overflow analysis
Closed-loop, task-level evaluation rather than per-frame accuracy alone
Consistent benchmarking discipline
Weaknesses
Evaluation is entirely synthetic
No statistical uncertainty anywhere
Absolute mesh accuracy is not contextualized
Incomplete training protocol and unclear baseline runtime
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
- All 5 strengths & 4 weaknesses, analyzed in detail
- 4 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is LEAP-NBV?
A lightweight framework that runs foundation-model-driven human mesh recovery and next-best-view (NBV) planning entirely on an edge device (NVIDIA Jetson Xavier NX), so a mobile robot can autonomously plan where to look next without cloud offloading.
How much does LEAP-NBV compress the original model?
It distills teachers of 327-687M parameters into a single 32M-parameter student, then quantizes to FP16: model size falls from 0.3-2.5 GB to 65 MB (30-60x smaller), and HMR inference drops from 257 ms to about 12 ms.
What accuracy does the distilled model lose?
On the paper's occluded synthetic benchmark, distillation recovers 6-7 mm of PA-MPVPE versus an undistilled 32M baseline (62.9 mm), with students near 56 mm; FP16 costs about 1 mm more, and closed-loop task metrics stay close to the teachers.
Is the evaluation based on real images?
No. According to the paper, all quantitative evaluation uses synthetic data: Blender-rendered DISC environments with SMPL-X bodies at 20-60% controlled occlusion; real AGORA images appear only in the training mix.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figures 1-6 are not analyzable from the extracted text (only captions); figure-derived numbers rely on caption text.
- The abstract appears truncated mid-sentence in the provided text.
- Training hyperparameters may exist in an appendix not included in the extraction.
- This is a v1 arXiv preprint (2609.23974); no peer review record is available.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.