“…Methods operating in a self-supervised manner aim to overcome this limitation by generating a training signal without relying on manual annotation. Instead they exploit information from other sensor modalities (Brooks and Iagnemma, 2012;Otsu et al, 2016;Castro et al, 2023;Seo et al, 2023;Higa et al, 2019;Meng et al, 2023;Zürn et al, 2021;Sathyamoorthy et al, 2022), or the interaction of the robot with the environment (Richter and Roy, 2017;Seo et al, 2022;Frey et al, 2023;Ahtiainen et al, 2017;Gasparino et al, 2022;Cai et al, 2022;Xue et al, 2023b;Sathyamoorthy et al, 2022;Cai et al, 2023;Jung et al, 2023). The generated supervision signal allows training a model that predicts a look-ahead estimate of the terrain, all without requiring the robot to be near to or interact with the terrain.…”