Welcome! We’re Justin and Matt, currently working on our MSc dissertations at Imperial College London.
Our research focuses on Vision-Language Action (VLA) models, with an emphasis on evaluating their performance in real-world settings using the SO-101 robotic arm.
We’ll be sharing regular updates on our progress, experiments, and findings as the project evolves.
Posts
-
Does VLA-0 Really Work? Reproducing NVIDIA's Results
I set out to reproduce the results reported in the VLA-0 paper, from the LIBERO benchmark to real-world robotics, to find out whether the model is as simple and effective as it claims. -
Why are VLAs so Complicated? Could the Simplest Solution Work?
A look at VLA-0, a surprisingly simple approach to robotic control that uses a standard vision-language model with no architectural changes, from NVIDIA Research. -
How Much Can You Change Before VLAs Break? Testing Environment Robustness
We tested a VLA by changing lighting, camera inputs, and the workspace. The result is strong robustness to visual changes, but surprising fragility to viewpoint and sensor shifts -
What Makes a Good VLA Dataset? Lessons from Real Experiments
After repeated failures, we shifted focus from models to data, empirically testing how dataset structure, consistency, and variance affect VLA performance. -
200 Episodes Later: Why More Data Didn’t Fix Our VLAs
We collected a larger 200 episode dataset with better visual setup, but still observed the same failures across all models, highlighting the importance of consistent actions over dataset scale. -
Are VLAs Overhyped? First Results on a Real Robot
Early experiments with VLAs on a real robot suggest that few-shot learning capabilities are overstated, with performance heavily dependent on data quality.
subscribe via RSS