Recti-Q: Feature-Space Rectification for OOD-Robust Quantized Perception in Edge Robotics

H. Yaghoubi*, P. Pilevar*, M. Lin (*equal contribution). IROS 2026.

  We identify and quantify the out-of-distribution (OOD) robustness gap introduced by 4-bit post-training quantization (PTQ) in large vision backbones deployed on resource-constrained robotic platforms. We then propose Recti-Q, a feature-space rectification method that trains a small frozen-backbone LoRA adapter using source-domain data only. Recti-Q preserves the memory savings of PTQ while closing most of the robustness gap under real-world sensor noise and weather corruption.

Recommended citation: H. Yaghoubi*, P. Pilevar*, M. Lin (*equal contribution). "Recti-Q: Feature-Space Rectification for OOD-Robust Quantized Perception in Edge Robotics." IROS 2026.