Recti-Q fixes quantized AI models' robustness for edge robotics

Robotic perception on edge devices often uses quantized neural networks (e.g., 4-bit PTQ) to fit within size, weight, and power constraints. While these models retain high accuracy on clean data, a new paper from Hamidreza Yaghoubi Araghi, Parastoo Pilevar, and Ming C. Lin reveals a critical flaw: they become brittle under real-world distribution shifts like sensor noise, bad weather, or unfamiliar environments. Across ImageNet-C and PACS benchmarks, 4-bit PTQ models show significant robustness degradation compared to their full-precision counterparts, despite negligible in-distribution accuracy loss.

To address this, the team proposes Recti-Q, a feature-space rectification framework that is architecture-agnostic (works on CNNs and Transformers) and requires no teacher model. It freezes the quantized backbone and trains a tiny LoRA adapter on the classifier head using only original source data. With less than 1% parameter overhead (as small as 6 KB), Recti-Q preserves over 99% of the memory savings from quantization while recovering most of the lost robustness—in some cases matching or exceeding FP32 performance.

Crucially, the minimal compute and memory footprint allow Recti-Q to be deployed as a low-bandwidth Over-The-Air (OTA) resilience patch for existing robot fleets. This means manufacturers can update perception models in the field without swapping hardware or retraining entire networks, making autonomous robots more reliable in unpredictable environments. The paper has been accepted at IROS 2026.