HSGNav helps home robots re-find objects that people have moved

Researchers led by Kaiyao Liao have presented HSGNav, a navigation method for assistive home robots that must find things from large furniture to small items on tables while people move objects around, in a paper posted on arXiv.

The method keeps lasting 3D entries for furniture apart from revisable guesses about objects on surfaces, which hold on to their visual evidence without needing their own geometry. The authors call this a Hypothesis Scene Graph. HSGNav chooses which kind of evidence to use from the size of the target, explores to see more surfaces and checks whether the target is there, dropping guesses that turn out wrong.

Because benchmarks such as GOAT-Bench rarely move objects in the middle of an episode, the team built FAROS-Bench (Find And Re-find Objects across Scales). HSGNav was competitive on GOAT-Bench and well ahead of earlier methods on FAROS-Bench, the authors report.