Gribincea, Sebastian (2026) Dreaming and Acting Beyond Boundaries: Model-Based Reinforcement Learning with Foundation Models. Master's Thesis / Essay, Artificial Intelligence.
|
Text
sebastiangribinceas4375548masterthesisfinal.pdf Download (12MB) | Preview |
Abstract
Foundation Models (FMs) trained on large datasets demonstrate strong representations across domains. Model-based reinforcement learning agents, such as DreamerV3, learn world models from raw observations but rely on encoders trained from scratch. This work studies whether pre-trained FMs can improve world model learning in a complex environment. A modular approach was developed to replace DreamerV3's encoder with pre-trained visual backbones, using Low-Rank Adaptation for parameter-efficient fine-tuning. Eight model variants were evaluated, combining two encoder architectures (CNN and Vision Transformer), two pre-training conditions, and two model capacities. All variants were trained on a wood collection task in Minecraft under identical conditions. Results show that the pre-training utility depends on encoder architecture. CNNs trained from scratch outperform randomly initialized ViTs, suggesting the advantage of inductive biases in data-limited settings. Pre-training recovers ViT performance, but degrades CNNs, suggesting domain-shift challenges when adapting natural image representations. Model capacity alone does not improve performance, but larger models benefit when combined with pre-training. The largest model using a pre-trained ViT achieved the highest performance. These findings suggest FM integration into model-based RL is feasible via PEFT methods, but effectiveness depends on the interaction between architectural priors and pre-trained representation quality.
| Item Type: | Thesis (Master's Thesis / Essay) |
|---|---|
| Supervisor name: | Sabatelli, M. and Cardenas Cartagena, J. D. |
| Degree programme: | Artificial Intelligence |
| Thesis type: | Master's Thesis / Essay |
| Language: | English |
| Date Deposited: | 03 Mar 2026 11:57 |
| Last Modified: | 03 Mar 2026 11:57 |
| URI: | https://fse.studenttheses.ub.rug.nl/id/eprint/37248 |
Actions (login required)
 |
View Item |