EmbodiedIntentBench scores drone mission software on whether missions succeed

Researchers led by Rajat Bhattacharjya have released EmbodiedIntentBench, a benchmark that judges how mobile robots adapt their computing at run time by whether their missions meet their goals, in a paper posted on arXiv.

The benchmark writes mission intent as structured contracts covering coverage, timing of evidence, reporting, urgency and resources, and runs simulated workloads for uncrewed aerial vehicles in which computing choices affect mission time, resource use and which observations count. The team compared fixed and adaptive perception strategies as urgency, connectivity and battery levels changed.

In a multi-phase depot study, a configuration with 87.5% target recall completed only 1 of 16 missions, while a configuration with lower recall completed 14. Replaying runs with Jetson AGX Orin CPU and GPU profiles changed which of the local and remote strategies did better, the authors report.