gnaroshi_vla
This workspace separates model adapters, methods, environments, and result locations so a VLA run can be inspected later.
Verified artifact
At a glance
- For
- Researchers comparing VLA implementations
- Primary use
- Configure a model and method combination, run a small check, and preserve the run context.
- Product status
- In development
- Platforms
- CLI
Research question
Which parts of a VLA experiment can be shared without erasing differences that matter?
The project explores how much experiment structure can be shared across VLA implementations without hiding model-specific behavior.
Approach
- Keep model-specific behavior behind explicit adapters.
- Separate reusable methods from environment and launch details.
Example workflow
Example experiment
Configure a small example run and inspect the evidence it leaves behind.
Choose the run
Select an architecture, method, environment, and experiment configuration.
Run a sanity check
Verify composition and launch wiring with a deliberately small check.
Inspect the manifest
Review the resolved configuration and environment snapshot.
Open the result directory
Follow the explicit result path without treating unchecked output as a conclusion.
The shown sanity run is example evidence, not a benchmark result.
Reproducibility
- Composed configuration records every selected layer.
- Run manifests and environment snapshots preserve context.
Current evidence
The current evidence establishes the experiment structure and a lightweight execution check only.
Unknowns
- Whether the same adapter boundary remains useful across more VLA families.
- Which efficiency comparisons remain fair across implementations.
Key features
- Separate model, method, environment, and experiment configuration
- Small adapter boundary for Seer and SimVLA
- Run manifests and environment snapshots
- Explicit result directories
Data and privacy
The public repository contains structure and lightweight evidence, not private datasets, credentials, or unpublished benchmark claims.
Tech stack
Languages
- Python
- Shell
Data
- YAML
Technology icons are repository-owned category symbols; the verified technology name remains visible.
Current state
The repository contains Seer and SimVLA integration structure, shared method modules, and run-context tools. It does not present benchmark results.
Current limitations
- The common adapter boundary is still being refined.
- Cross-model result comparisons need reviewed runs before publication.
Technical notes
- Python owns orchestration and adapters.
- YAML keeps model, method, environment, and node choices explicit.
Next milestones
- Complete consistent sanity checks across adapters.
- Record reviewed comparison runs without overstating their scope.