Dbp-dev12 / vla-preflight

# VLA Preflight Workbench
**Move from robot-learning data checks to reproducible training and supervised rollout evidence.**
中文 · Workflow guide · Validation · Audit checks
v0.6.0 alpha keeps the laptop-only path and adds a careful path for users with CUDA GPUs and physical robots. It diagnoses compute without leaking host identity, generates safety-gated LeRobot bring-up plans, and reports real or simulated rollout success with uncertainty. It never silently connects to hardware or runs a robot.
| Capability | What actually happens |
|---|---|
| Inspect | Audit schema, numeric trajectories, statistics, sampler traces and video references; view images and action curves |
| Prepare | Group exact numeric duplicates, split whole episodes, fit **train-only** normalization, fingerprint inputs |
| Train | Run a randomly initialized image + byte-text + state model with masked action chunks and held-out evaluation |
| Compare | Record baselines, loss curves, last/best checkpoints, sampler traces and compatible resumed runs |
| Export | Filter episodes into a new v3 dataset, reindex and recompute numeric statistics |
| Extend | Check a LeRobot 0.6 environment and generate a SmolVLA plan with a separate training split |
| Diagnose | Report CUDA devices, VRAM, optional packages and LeRobot commands without hostnames, usernames or paths |
| Bring up | Validate a declared robot/camera/calibration setup and generate reviewable teleoperation/recording commands |
| Evaluate | Summarize rollout success, interventions, failure modes and Wilson confidence intervals by task/checkpoint |
Tiny VLA is a reference trainer, **not a pretrained foundation model**. Offline action error is not robot task success. The LeRobot bridge and artifact capture are regression-tested, but this release does not claim a live SmolVLA/CUDA result. This project does not replace LeRobot.
## Start locally
Use Python 3.10–3.13. Create and activate an environment:
 [code] 
Open `http://127.0.0.1:8765`. Inspect episodes → create preparation bundle → select it → train. The default 24-episode synthetic set can overfit: compare with the constant-action baselines before interpreting decreasing training loss.
After installation, Windows users can run `Run-Workbench.cmd` to generate and open the demo. CPU wheels avoid unnecessary CUDA downloads. `uv sync --extra dev --locked` installs core/development dependencies; the lockfile's optional train extra uses the default package index and may bring large accelerator dependencies on Linux.
## CLI workflow
 [code] 
Outputs must be outside source data. Preparation, training, export and plan destinations must be new directories. Resume uses a new run directory; `--steps` is the **total target**. Source fingerprints and preparation artifacts are checked before training.
Only need preflight? `pip install -e .` installs no PyTorch/imaging dependencies:
 [code] 
The final command intentionally exits 1 and detects `WINDOW_CROSS_EPISODE`. Audit exit codes: 0=no er