Skip to main content

​
env run

Executes a project’s mission phase-by-phase, logging every tick and running the LLM judge at the end for a verdict. 
cadenza env run rescue-dog --headless
env run requires the gym extra (MuJoCo). Install with pip install -e ".[gym]".

​
Flags

FlagDefaultDescription
--headlessviewer shownRun without the MuJoCo viewer. Works on any Python (no mjpython needed). Use this in scripts/CI.
--policy <p>scriptedDriver. scripted parses each phase’s goal_prompt into actions. lora drives with your fine-tuned LoRA action head (needs the lora extra).
--xml <path>project defaultOverride the MuJoCo scene XML.
cadenza env run rescue-dog --headless --policy lora

​
Output

Each run writes two files under <project>/.cadenza-env/:
FileContents
<run-id>.log.jsonlOne record per executed action: phase, reward, observation summary.
<run-id>.finetune.jsonl(prompt, action, reward, …) rows rendered via vla_finetune.prompt_template.
The scripted driver exercises the full loop and emits fine-tune records but is not a learned policy. To drive the same mission with your own model, see Custom environments.

​
env stats

Aggregate motion and reward analytics across a project’s cached runs. 
cadenza env stats rescue-dog
Reports per-phase success, reward breakdowns, and motion summaries, so you can see whether edits to your reward shaping or policy actually helped.

​
env cache

Show or prune the per-project .cadenza-env/ run directory to manage disk. 
cadenza env cache rescue-dog                    # show what's cached
cadenza env cache rescue-dog --keep-last 5      # keep only the 5 newest runs
cadenza env cache rescue-dog --older-than 7d    # drop runs older than 7 days
cadenza env cache rescue-dog --clear            # remove all cached runs
FlagDescription
--keep-last <n>Keep the n most recent runs, delete the rest.
--older-than <dur>Delete runs older than a duration (e.g. 7d, 24h).
--clearDelete all cached runs for the project.