Inference template

simulo create mybot --type inference

A full policy lifecycle in one self-contained app: train produces its own checkpoint, so evaluate and rollout are never stranded waiting on another app to have run first. All three are @app.jobs in the same file, sharing two volumes — one for checkpoints, one for reports.

Shape

checkpoints = simulo.Volume.from_name("mybot-checkpoints", create_if_missing=True)
reports = simulo.Volume.from_name("mybot-reports", create_if_missing=True)

app = simulo.App("mybot", mounts={"/checkpoints": checkpoints, "/reports": reports})

@app.job(system=simulo.SystemType.TIER_1, timeout=8 * 60 * 60, retries=2, callbacks=[simulo.callbacks.ResumableCheckpoint(every=50)])
def train(num_envs: int = 512, max_iterations: int = 20) -> dict:
    ...  # saves BOTH the trainer checkpoint AND an exported TorchScript policy

@app.job(system=simulo.SystemType.TIER_1, timeout=2 * 60 * 60)
def evaluate(num_episodes: int = 10, num_rounds: int = 5) -> dict:
    ...  # replays the checkpoint over several rounds, aggregates with numpy

@app.job(system=simulo.SystemType.TIER_1, timeout=1 * 60 * 60)
def rollout(num_steps: int = 200) -> dict:
    ...  # plays the exported policy with simulo.RLPlayer, records an MCAP

No @app.entrypoint — with three jobs and no entrypoint, --job NAME picks which one simulo run submits, and the remaining flags map onto that job’s own parameters (see App & Jobs). Only one stage is submitted per simulo run; omit --job and it refuses, naming the three jobs rather than guessing. After a submit, simulo run names the other jobs it didn’t run in one line (e.g. “Other jobs in app.py: evaluate, rollout (submit with: simulo run app.py –job NAME)”) — a pointer to the flag, not a per-stage copy-pasteable command with its own flags filled in.

Run each stage, in order

evaluate and rollout both need train’s checkpoint volume populated first — running them before train fails with a clear “checkpoint not found” error, not a silent no-op.

1. train

simulo run app.py --job train --num-envs 512 --max-iterations 20

Trains a PPO policy and saves two outputs to the checkpoint volume: the trainer checkpoint evaluate reads, and an exported TorchScript policy rollout plays directly (trainer-free).

2. evaluate

simulo run app.py --job evaluate --num-episodes 10 --num-rounds 5

Replays the checkpoint over several independent rounds and aggregates mean/std reward and a success rate against a configurable reward target, writing a JSON report to the reports volume.

3. rollout

simulo run app.py --job rollout --num-steps 200

Plays the exported policy with simulo.RLPlayer (no trainer) and records the run to an MCAP flight recording — see Record and inspect MCAP.

Where to edit

Same task-authoring surface as the training template (MybotTask.build() / get_observations() / get_rewards() / get_dones() / apply_actions() / reset_idx()), plus:

• evaluate — the metrics you aggregate and the reward target you score against (the scaffold reads it with os.environ.get(...), with a built-in default).
• rollout — recording configuration (simulo.RecordConfig) and how many steps to play.

Observe

simulo logs --follow      # each stage, while it runs
simulo result             # each stage's own return value
simulo recordings         # after rollout — downloads the MCAP