Electronic Thesis/Dissertation
 

Improving Success Rate in Robotics Task Completion Using Model-Based Reinforcement Learning

Open Access Deposited

(ii) a multimodal fusion, MambaBind, which employs Mamba state-space models and a contrastive alignment objective to better combined multiple sensor inputs

computational efficiency, multimodal sensor fusion, and multi-agent scalability. Reliance on traditional recurrent neural networks can bottleneck learning in long horizon/sparse reward environments, simple concatena- tion of multiple sensor inputs underutilizes cross-modal structure, and the limited availability of standardized model-based multi-agent reinforcement learning algorithms particularly public multi- agent variants of DreamerV3 further complicates adoption. This praxis develops and evaluates a hybrid MBRL framework built upon DreamerV3 with three core innovations

(i) integration of the Light Recurrent Unit and the Stack Recurrent Cell to improve computational efficiency and long-term dependency modeling

and (iii) a scalable multi-agent design that uses variational autoencoders to enable a shared world model to generalize across scenarios with varying numbers of agents. The framework is evalu- ated on Atari-100k, Crafter, CARLA, and SMACv2. Results show that the recurrent variants are more performant than both baseline on long-horizon/sparse reward tasks. MambaBind improves performance over concatenation in CARLA. The multi-agent framework scales and generalizes, outperforming established model-based and model-free multi-agent baselines. Overall, this work provides a hybrid scalable framework that improves sample efficiency, accelerates training, and enables more capable general-purpose agents for complex robotics applications.

Model-based reinforcement learning is promising for robotics, yet leading approaches suchas DreamerV3 face three problems that hinder real-world deployment

Author Language Date created Type of Work License
  • All rights reserved
Rights statement GW Unit Degree Advisor Committee Member(s) Persistent URL

Notice to Authors

If you are the author of this work and you have any questions about the information on this page, please use the Contact form to get in touch with us.

Thumbnail Title Date Uploaded Visibility Actions
Preview of Nguyen_gwu_0075A_17718.pdf Nguyen_gwu_0075A_17718.pdf 2025-12-12 Open Access