Dataset structure: dual-arm manipulation demonstrations collected from Mercury X1 teleoperation.
Dataset tasks include manipulating fruits and objects such as banana and grapes.
The dataset is formatted into observations, language instructions, and robot action trajectories for VLA training.
Preprocessing before finetuning aligns camera frames, action chunks, and language commands.
OpenVLA-OFT finetuning uses LoRA adapters with the collected robot demonstration dataset.
Hyperparameters are present in the thesis for fine-tuning runs, including learning rate, batch size, training steps, and LoRA rank.
Dataset sizes and sampling rates mentioned in the thesis include 480 teleoperation demonstration episodes, a 50 Hz robot control loop, and 25 Hz synchronized camera sampling.
GPU memory and VRAM numbers mentioned include 96 GB for high-memory training, 62 GB observed peak usage, 24 GB for a smaller available GPU configuration, and 16 GB for deployment/inference constraints.
Compared models are OpenVLA, finetuned OpenVLA-OFT, and AgenticVLA.
Experiments and results compare standard VLA behavior against AgenticVLA on semantic relationship, prompt complexity, and self-awareness benchmarks.
Limitations include sim-to-real transfer, dataset scale, hardware constraints, and robustness in open environments.