Wall-clock latency breaks synchronous navigation benchmarks
Most work on finding an unseen object in an unknown environment is developed in synchronous simulators where the environment waits for the agent and inference time costs the task nothing. A preprint on arXiv dated 29 August shows that current zero-shot object-navigation methods degrade consistently once wall-clock time counts against the budget. Their RTNav architecture treats inference latency, asynchronous environment stepping and bounded compute as explicit design constraints. On real-time versions of the HM3D-v1, HM3D-v2 and HM3D-OVON benchmarks it raises the success rate by up to 11 points and success weighted by completion time by up to 5.1 points. The results were obtained in simulated environments. The work is a preprint that has not been peer reviewed.[1]