Hadrien Crassous, Mohamed Yassine Kabouri, Minahil Raza +2 more
A robot arriving in an unfamiliar building has an object detector that was trained somewhere else and will make mistakes here. Improving it means gathering examples and having them labelled, and both cost something: driving around consumes time, labelling consumes human effort.
Standard active learning assumes a pool of candidate images sitting ready to be chosen from. An embodied agent has no such pool. It has to physically go and collect the candidates, so the choice of what to label is entangled with the choice of where to drive, and a promising sample on the far side of the building may not be worth the trip.
The approach targets the detector's failure cases specifically, and both budgets are respected explicitly rather than one being treated as free. That is the honest formulation of the problem a real robot faces.
This paper studies how to adapt a computer vision object detector to an unknown environment under both a robot navigation time and annotation budget constraint. Our approach selects informative robot trajectories and image samples to retrain the detector, explicitly targeting its failure cases. Formally, the approach is an embodied variant of batch active learning, where at each round an agent has a limited navigation budget to collect candidate samples and a limited…
MAGiSt3R: Multi-Agent Feed-forward 3D Reconstruction from Monocular RGB Videos
arXiv (cs.CL) · July 16, 2026Expanding the Lexicon of Ge'ez Based African Languages: A Comparative Study of Amharic and Tigrinya
arXiv (cs.LG) · July 16, 2026Delocalization of bias in unadjusted Hamiltonian Monte Carlo and underdamped Langevin
arXiv (cs.LG) · July 16, 2026BadWAM: When World-Action Models Dream Right but Act Wrong