Track papers across labs, topics, and trends.
Search papers, labs, and topics across Lattice.
This paper addresses the challenge of quantization-induced robustness degradation in robotic perception systems, particularly under distribution shifts like sensor noise and severe weather. The authors introduce Recti-Q, a lightweight feature-space rectification framework that utilizes a small classifier-head LoRA adapter to recover lost robustness in 4-bit post-training quantized models while maintaining in-distribution accuracy. The results demonstrate that Recti-Q can match or exceed the performance of full-precision models with minimal parameter overhead, making it a practical solution for edge robotics applications.
Recti-Q recovers lost robustness in quantized models, achieving performance comparable to full precision while preserving over 99% of memory savings.
Robotic perception pipelines increasingly rely on large vision backbones deployed on SWaP-constrained edge platforms, making post-training quantization (PTQ) attractive for real-time inference. However, while PTQ often preserves clean in-distribution accuracy, we show that it can substantially degrade reliability under deployment-relevant distribution shifts (e.g., sensor noise, severe weather, and novel operating environments), creating a Quantization-Induced Robustness Gap. Across foundational vision benchmarks (ImageNet-C and PACS), 4-bit PTQ models exhibit pronounced robustness degradation despite negligible ID accuracy loss. To address this, we propose Recti-Q, a lightweight feature-space rectification framework that freezes the quantized backbone and trains a small classifier-head LoRA adapter using only source data. Recti-Q is architecture-agnostic across CNNs and Transformers, supports efficient teacher-free training, and recovers a significant portion of the lost robustness, in some cases matching or exceeding FP32 performance. At less than 1% parameter overhead (as small as 6 KB), Recti-Q preserves over 99% of PTQ memory savings, adds negligible compute, and enables low-bandwidth Over-The-Air (OTA) resilience patching for deployed robotic fleets operating in unpredictable physical environments.