Title: Towards Connected Intelligence: Empowering Robotic Applications with Agentic AI Frameworks
Abstract: The thesis studies Agentic AI frameworks integrated with Vision-Language-Action models for embodied robotic systems.
AgenticVLA end-to-end robot manipulation pipeline is the proposed system for the Mercury X1 humanoid robot.
OpenVLA foundational vision language action model is used as the baseline VLA.
OpenVLA is a 7B-parameter open-source Vision-Language-Action model.
OpenVLA-OFT fine-tuned with on-robot data using LoRA adapters is described as the optimized model.
OpenVLA-OFT uses parallel decoding and action chunking and achieves 43x faster throughput than base OpenVLA.
AutoGen multi-agent framework orchestrates AgenticVLA skills for planning, task reasoning, and tool coordination.
Research questions ask how agentic frameworks improve VLA robotic performance and how embodied cognition supports connected intelligence.
Main objectives include teleoperation, data collection, training the VLA, and Agentic AI integration.
Agentic AI components in the thesis include perception, planning, and action execution.