PaTSy-Neural-EM: Geometry-Aware Truth Discovery for Real-Time Multi-Agent Perception

ICLR 2026 Conference Submission21711 Authors

Keywords: Multi-agent Perception, Truth discovery, Sensor Fusion, Expectation-maximization(EM), V2X, OPV2V, Confusion matrices, Neural reliability head, Scene geometry features, Trust-region optimization

TL;DR: PaTSy-Neural-EM combines EM with neural reliability prediction to assess agent trustworthiness via geometric context, achieving robust real-time multi-agent perception under occlusion and long-range conditions.

Abstract:

We revisit truth discovery (TD) for multi-agent perception and present PaTSy-
Neural-EM, a geometry-aware EM framework that learns state-conditioned reliability
while preserving the interpretability of Dawid–Skene (DS) confusion
matrices. In dynamic V2X scenes, reliability varies with range, incidence angle,
occlusion, latency, and agent identity. PaTSy injects this context via a log-linear
reliability head whose outputs additively correct DS logits and are renormalized
with a softmax to yield valid context-dependent confusion columns. To stabilize
joint learning, we introduce a gentle-Π schedule: (i) warm-start Π with DS, (ii)
freeze Π while training the head, and (iii) unfreeze with a KL trust-region tether.
We further add physics-inspired regularizers: range-monotonicity and angular
smoothness. The resulting model runs in real time, remains DS-compatible, and
yields better calibration and hard-slice robustness at DS-level top-1 accuracy
on V2X-Real. On OPV2V under zero calibration, our best run improves over DS
by +0.9% absolute on both validation and test.

Primary Area: unsupervised, self-supervised, semi-supervised, and supervised representation learning

Submission Number: 21711