Skip to main content
This walks you from a fresh install to a fully run, scored mission with fine-tuning data on disk. It assumes you’ve installed cadenza and have the gym extra (or are happy to stop before env run).
1

Sign in

Your administrator issues you a token . Sign in once:
Verify:
See Authentication for details. Sign-in is required for env action commands and for any governed command that returns a verdict (env lora eval, env residual train/eval/bench, env distill eval, env vla grd/eval). Scaffolding, running, stats, and data export all work signed-out.
2

Scaffold a project

Creates ./rescue-dog/ with an env.json mission spec, a main.py driver, requirements.txt, and a README. Inspect the mission:
This renders the scene zones, objects, and the ordered mission phases.
3

Run the mission

Executes the mission phase-by-phase with the built-in scripted driver, logs every tick, and automatically runs the LLM judge at the end for a verdict. Drop --headless to watch it in the MuJoCo viewer.
4

Inspect results & make training data

Each run writes to rescue-dog/.cadenza-env/:
  • <run-id>.log.jsonl: one record per action (phase, reward, observation).
  • <run-id>.finetune.jsonl: (prompt, action, reward) records for VLA training.
Analyze and export:

​
What next

Author a mission

The full env.json schema: phases, predicates, reward shaping.

Build actions

Create and sync reusable robot actions to your account.

Fine-tune

env train and env lora to improve a policy.

Drive it from Python

Run the same mission with your own VLA via the SDK.