Podrobnosti studentského projektu

Seznam
Téma:Strojové učení pro robotickou exploraci
Katedra:Katedra kybernetiky
Vedoucí:MSc. Ruslan Agishev
Vypsáno jako:Bakalářská práce, Semestrální projekt
Popis:Goal: correction of sensory measurements that are input to SLAM (Simultaneous Localization and Mapping) system.
Sensory data: RGB-D (depth) images.
SLAM system: GradSLAM, differentiable implementation of PointFusion (ICP-SLAM).

The core of the depth correction (completion) pipeline is the differentiable SLAM module, which takes as input RGB-D images and outputs camera trajectory and map estimate. However, the input sensory data could have noise and missing depth values.

Therefore it is important to introduce a Depth Correction (DC) module, before the propagation of the depth measurements through the SLAM module. The DC module could also take as input another data (normals).

DC and ∇SLAM are differentiable modules, which should allow propagating gradients through the entire pipeline from input sensory data to the resultant map and camera trajectory. Once we have a constructed map (we can use ground truth trajectory if it is available), we can calculate the loss:
(1) reconstruction error (for example Chamfer distance between constructed and truth point clouds) using the ground truth map (available in datasets or in Subt simulator),
(2) localization error (using the ground truth poses from datasets or the simulator).

Once the loss is computed, the weights of the DC module could be updated (iterative process). We run the optimization process (iterations) until satisfied with the reconstruction error.
Literatura:[1] K. Murthy J., et al, ∇SLAM: Automagically differentiable SLAM, ICRA, 2020

[2] P. Karkus, et al, Differentiable SLAM-net: Learning Particle SLAM for Visual Navigation, CVPR, 2021

[3] D. S. Chaplot, et al, Active Neural Localization, ICLR, 2018

[4] S. K. Gottipati, et al, Deep Active Localization, RAL, 2019

[5] Thomas M Howard and Alonzo Kelly, Optimal rough terrain trajectory generation for wheeled mobile robots, The International Journal of Robotics Research, 2007

[6] Vojtěch Šalanský, et al, Pose consistency kkt-loss for weakly supervised learning of robot-terrain interaction model, RAL, 2021

[7] Zichao Zhang and Davide Scaramuzza, Fisher information field: an efficient and differentiable map for perception-aware planning, arXivpreprint arXiv:2008.03324, 2020.
Za obsah zodpovídá: Petr Pošík