Joshua Adrian Cahyono

• Developed and enhanced state-of-the-art diffusion planners for robotics autonomous navigation in MuJoCo simulation environments and ROS Gazebo.
• Conducted analysis of 3 different algorithms (behavior cloning with CNN, imitation learning with Transformer and Diffusion Planner) using PyTorch.
• Tested and deployed algorithm to physical mobile robot using ROS.

Co-developed "Otter," a Visual Language Model inspired by Google DeepMind's Flamingo, with refined training and evaluation pipelines (publication in review).

Curated and integrated 5 visual-text datasets and evaluation pipelines.

Conducted a comparative analysis of Otter, IDEFICS, and LLAMA2 models.

Co-developed LMMS-Eval pipeline and co-authored a publication on its methodology.

• Collaborated with research scientists to develop an AI-driven solution for optimizing diabetes treatment using Reinforcement Learning.
• Transformed the MIMIC-III dataset from Google BigQuery to PostgreSQL and DataFrame formats and built a robust evaluation pipeline to enable efficient testing of algorithms such as DQN, DRQN, DRQN-EWC, and MoE.
• Designed and implemented a software IP for a customized Offline Reinforcement Learning Environment Framework, based on OpenAI's Gymnasium framework tailored to medical dataset.
• Publication in review.