A robotics preprint reports that PVRA, a robot-vision framework, reached a Step Acc@0.15d score of 0.893 on 280 evaluated samples from simulated Nema17 progressive-assembly scenes. The study’s authors interpret that score as 89.29% of predictions meeting a roughly 7-to-8-millimetre tolerance for an average 51-millimetre Nema17 object. PVRA’s score was higher than CAD-ICP-PCA with ground-truth masks, CAD-ICP-PCA with PVRA-predicted masks and FoundationPose with PVRA-predicted masks, but lower than FoundationPose with ground-truth masks, which scored 0.989.
PVRA takes RGB-D point clouds, combining colour and depth information, and predicts whether each point belongs to the target, the base or the background. It then estimates a six-degree-of-freedom pose for the target object and a six-degree-of-freedom assembly pose. The training jointly supervises segmentation and two offset predictions, using focal loss for segmentation and offset loss for offset prediction.
The comparison depended on the input masks
The evaluation used 431 assembly instances involving five objects across four assembly steps, generating 8,620 instances in total. The data were split into 60% for training, 20% for validation and 20% for testing.
PVRA was compared with two CAD-based pose baselines: CAD-ICP-PCA and FoundationPose. Both were evaluated with annotation masks and with masks predicted by PVRA; the baselines used known base-target transformations and assumed ideal segmentation.
The main pose measure, Step Acc@0.15d, records how often an estimate falls within a fixed error threshold. The study also reports Target AUC, Assembly AUC and SLA-AUC, which summarize accuracy across thresholds from 0 to 0.50d, plus scores at 0.05d, 0.10d and 0.20d.
A strong result, with an important qualification
On the 280 samples used for its evaluation, PVRA recorded an SSA role-segmentation score of 0.833, a Target AUC of 0.745, an Assembly AUC of 0.790 and an SLA-AUC of 0.757. Its Step Acc@0.15d was 0.893.
At that threshold, PVRA’s 0.893 exceeded CAD-ICP-PCA with ground-truth masks at 0.344, CAD-ICP-PCA with PVRA-predicted masks at 0.286 and FoundationPose with PVRA-predicted masks at 0.771. It was below FoundationPose with ground-truth masks, which scored 0.989.
FoundationPose’s reported performance varied by mask condition. The discussion describes it as highly accurate with ground-truth masks but lower with sparse PVRA-predicted masks. FoundationPose with PVRA-predicted masks processed 53 fewer samples than PVRA, and the authors describe PVRA as more consistent across thresholds in that comparison.
Direct comparisons require caution because the methods did not all have the same number of usable test samples. PVRA can fail during keypoint voting, CAD-ICP-PCA can fail when too few segmented points are available, and FoundationPose can reject small masks. The paper reports no confidence intervals, hypothesis tests or other uncertainty estimates for these results.
What the test leaves open
All of the evidence comes from simulation. The authors identify sensor noise, calibration errors, reflections and lighting conditions as real-world challenges, and describe transfer to real settings as future work.
The formulation assumes a fixed assembly sequence, an explicitly defined target with a six-degree-of-freedom pose at each step, gravity-stable objects and only one object in contact with the target—the base object. These assumptions leave open whether the framework can handle changing sequences, multiple contacts or unstable configurations.
PVRA produces pose outputs for downstream manipulation, but it is not a complete controller. The reported scores therefore describe a synthetic pose-perception evaluation under the study’s fixed assumptions.
Paper data and sources
Original title: PVRA: A Pointwise Key-point Voting Framework for Robotic Assembly
Authors: Kulunu Samarawickrama, Roel Pieters
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-20
DOI: Not available
Original paper · Full text