Research | Gnaroshi

01

What kind of experiment structure makes different VLA models comparable on common ground?

Different VLA repositories organize models, environments, and results in different ways. That makes the same experiment difficult to run and inspect across models.

What I’m testing: I’m testing whether a shared run structure can make comparisons clearer while small adapters keep model-specific behavior visible.
What remains: The shared structure still needs to preserve details that affect correctness.

gnaroshi_vla

Generated illustration of a robot arm using a camera and a visual instruction to reach for a red block among several objects. — **Generated concept scene**An illustration of the research question, not experiment evidence.

02

How can computation be reduced without hiding changes in model behavior?

Saving computation matters only when the resulting behavior can still be checked. Reused state and approximations need clear boundaries.

What I’m testing: I’m looking at explicit reuse and update paths that can be measured across experiments.
What remains: The useful boundary may change with the model, observation history, and action decoder.

Experiment setup

Diagram comparing full recomputation with a smaller update path that reuses cached model state. — **Technical diagram**An original explanatory diagram for this research question.

03

How can AI tools support research without replacing the researcher’s judgment?

Generated summaries and scores can save time, but they can also hide uncertainty. The original question and the researcher’s own check still need to remain visible.

What I’m testing: I’m exploring tools that help organize notes, suggest questions, and make revisiting easier without deciding what is true on the researcher’s behalf.
What remains: The right amount of structure will only become clear through repeated use.

Writing Reading

Diagram of a research loop: read, explain from memory, check what was missed, revisit, then implement or write. — **Technical diagram**An original explanatory diagram for this research question.