Training template

simulo create mybot --type training     # --type training is the default

The simplest starter: one job that trains a PPO policy to balance a pole on a cart, and saves the trained checkpoint to a durable volume.

Shape

cartpole = simulo.Asset.from_registry("simulo/robot/cartpole:v1")

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

app = simulo.App("mybot", mounts={"/out": vol})  # runs on the default Simulo runtime

with app.runtime.imports():
    import torch  # resolved only on the worker, never at submit

class MybotTask(simulo.Task):
    ...  # build(), get_observations(), get_rewards(), get_dones(), apply_actions(), reset_idx()

@app.job(
    system=simulo.SystemType.TIER_1,
    timeout=8 * 60 * 60,
    retries=2,
    callbacks=[simulo.callbacks.ResumableCheckpoint(every=50)],
)
def train(num_envs: int = 4096, max_iterations: int = 200) -> dict:
    env = simulo.LearningEnv(task=MybotTask(), num_envs=num_envs, device="cuda", ...)
    trainer = simulo.RLTrainer(env=env, algorithm="PPO", device="cuda", seed=42)
    stats = trainer.train(max_iterations=max_iterations)
    trainer.save(f"{vol.path}/mybot_final.pt")
    ...
    return {"checkpoint": ..., "num_envs": num_envs, **stats}

No @app.entrypoint — train is the app’s only job, so simulo run maps --num-envs/--max-iterations straight onto its own parameters.

Where to edit

• MybotTask — build() constructs the scene (swap the pinned global cartpole ref for your own published asset; see Bring your own robot). get_observations() / get_rewards() / get_dones() / apply_actions() / reset_idx() are your task’s actual learning logic — observation vector, reward shaping, termination condition, and how actions map onto the robot’s joints.
• train — training hyperparameters (num_envs, max_iterations), resources (system=, timeout=), and what the job returns. Its parameters are the app’s CLI flags — simulo run app.py --help always reflects this function’s current signature.

retries=2 plus ResumableCheckpoint(every=50) mean a retried or preempted run resumes from its latest checkpoint instead of restarting — see Continue or resume for exactly what that callback declares and how it differs from simulo run --from.

Run it

simulo run app.py --num-envs 4096 --max-iterations 200

--num-envs and --max-iterations are this app’s flags, mapped straight onto train’s own parameters (no @app.entrypoint needed) — not flags simulo run itself defines. See Run & submit for the full distinction.

simulo logs --follow      # watch it train
simulo result             # best_reward, checkpoint, iterations
simulo models             # best.pt / latest.pt, ready to download