| Meno: | Miroslav |
|---|---|
| Priezvisko: | Krupa |
| Názov: | Self-Supervised Robotic Trajectory Planning with Obstacle Avoidance |
| Vedúci: | RNDr. Kristína Malinovská, PhD. |
| Rok: | 2026 |
| Kµúčové slová: | robotic trajectory planning, self-supervised learning, obstacle avoidance |
| Abstrakt: | This bachelor's thesis focuses on self-supervised trajectory planning for robotic manipulators in simulated environments with obstacles using recurrent neural networks. As part of the thesis, a dataset of robot trajectories and transitions was generated using a robotic manipulator in a simulated environment. Subsequently, forward and inverse models were trained to approximate the system dynamics and the actions between individual environmental states. These models were then used as supervisory mechanisms during the training of recurrent planners. The thesis compares several trajectory model architectures and training approaches, including training with geometric priors, supervised pretraining, and various combinations of supervisory models aimed at limiting exploitation of the learning feedback mechanism. Experimental results show that the proposed approach is capable of generating geometrically consistent and feasible trajectories. At the same time, it was demonstrated that geometric regularization and pretraining significantly improve the stability of the generated trajectories. The thesis further analyses error propagation, trajectory feasibility, and the inference time requirements of the individual models. The obtained results suggest that self-supervised recurrent trajectory planning represents a promising and computationally efficient approach to robot motion planning in obstacle-containing environments. |
Súbory bakalárskej práce:
| bc_thesis_source_code.zip |
| bc_thesis-Krupa.pdf |
Súbory prezentácie na obhajobe: