Perception router cuts robot deadline misses by up to 9.4%

A new arXiv preprint (arXiv:2608.00322) from Sparsh Roy, Vihan Aggarwal, and Davin Yin tackles a double problem in real-time robot perception: cameras degrade in rain, mud, and fog, while the shared onboard processor delivers inconsistent compute to perception due to planning and control tasks. Existing systems model these two stressors separately. The authors propose a perception router that maintains probabilistic estimates of both sensor-fault state and compute-contention state, couples them with a noisy-OR term, and uses the joint belief to pick one of four detector configurations—YOLO11x or YOLO11n at 1280 or 640 pixels—so each frame finishes before its deadline.

In experiments with co-occurring stressors, the coupled policy cuts the deadline-miss rate by 1.1 to 9.4 percentage points against a policy that treats the stressors independently, with a sign-test p = 0.001 across 10 sequences and 6 conditions. Routing overhead is just tens of microseconds per frame. However, the authors honestly report a null result: across eight real RADIATE adverse-weather sequences and three workload proxies, after Benjamini-Hochberg correction and replication, none of 24 tests found natural coupling. They scope the routing result as a proof of mechanism, and the released evaluation pipeline lets deployments test for coupling on their own traces.