Multi-simulated deep reinforcement learning based conversational robot arm/
Bumagat, Kyle Mhiron P.
Multi-simulated deep reinforcement learning based conversational robot arm/ Kyle Mhiron P. Bumagat, Arianne Joy D. Evangelista, John Louis D. Lagramada, Jhon Mark G. Remollo, Nicamae G. Tamundong, and Althea R. Villanueva.-- - Manila: Technological University of the Philippines, 2025. - x, 90pages: 29cm.
Bachelor's thesis
College Of Engineering.--
Includes bibliographic references and index.
Adaptive control that dynamically responds to perturbations is a difficult problem
to solve using classical techniques in robot manipulation. This raises a critical issue in
creating a fully collaborative robotic system. In this study, deep reinforcement learning
was used to train a robot arm in reaching a specific 3D point in space using 4096 parallel
environments. The rewards were shaped to minimize action rate and velocity under a
curriculum with domain randomization. This study simulated, gain-tuned, and trained the
agents using the Isaac Lab framework. Within 6.24 minutes, a performant model arrived,
essentially compressing months of training to minutes. Proxima policy optimization (PPO)
was used with a multi-layer perceptron (MLP) neural network backbone. A modular ROS
2 package was also presented for bridging natural language understanding to physical
systems. The package employs both cloud and edge computing of state-of-the-art speech-
to-text, text-to-speech, and large language models. The developed system was evaluated
for its educational impact using a structured learning module. Results showed significant
improvements between the pre-test and post-test, with a t-value of -19.329 and p < .001.
The mean scores for the evaluated factors: Attention, Relevance, Confidence, and
Satisfaction ranged from approximately 3.9 to 4.3, with Cronbach’s alpha values indicating
high reliability. The Structural Equation Model (SEM) indicated a good fit, SRMR = 0.071.
Future enhancements should focus on increasing sample size, adding beginner-friendly
resources, and expanding advanced ROS 2 topics. Further development of the
conversational robot arm with interactive features can also improve student engagement
and learning.
Robot manipulation
Adaptive control
Deep learning
BTH TK 870 / B86 2025
Multi-simulated deep reinforcement learning based conversational robot arm/ Kyle Mhiron P. Bumagat, Arianne Joy D. Evangelista, John Louis D. Lagramada, Jhon Mark G. Remollo, Nicamae G. Tamundong, and Althea R. Villanueva.-- - Manila: Technological University of the Philippines, 2025. - x, 90pages: 29cm.
Bachelor's thesis
College Of Engineering.--
Includes bibliographic references and index.
Adaptive control that dynamically responds to perturbations is a difficult problem
to solve using classical techniques in robot manipulation. This raises a critical issue in
creating a fully collaborative robotic system. In this study, deep reinforcement learning
was used to train a robot arm in reaching a specific 3D point in space using 4096 parallel
environments. The rewards were shaped to minimize action rate and velocity under a
curriculum with domain randomization. This study simulated, gain-tuned, and trained the
agents using the Isaac Lab framework. Within 6.24 minutes, a performant model arrived,
essentially compressing months of training to minutes. Proxima policy optimization (PPO)
was used with a multi-layer perceptron (MLP) neural network backbone. A modular ROS
2 package was also presented for bridging natural language understanding to physical
systems. The package employs both cloud and edge computing of state-of-the-art speech-
to-text, text-to-speech, and large language models. The developed system was evaluated
for its educational impact using a structured learning module. Results showed significant
improvements between the pre-test and post-test, with a t-value of -19.329 and p < .001.
The mean scores for the evaluated factors: Attention, Relevance, Confidence, and
Satisfaction ranged from approximately 3.9 to 4.3, with Cronbach’s alpha values indicating
high reliability. The Structural Equation Model (SEM) indicated a good fit, SRMR = 0.071.
Future enhancements should focus on increasing sample size, adding beginner-friendly
resources, and expanding advanced ROS 2 topics. Further development of the
conversational robot arm with interactive features can also improve student engagement
and learning.
Robot manipulation
Adaptive control
Deep learning
BTH TK 870 / B86 2025