Templates
simulo create <name> --type <type> scaffolds a runnable app folder from one
of three starter templates, bundled inside the simulo package itself —
no network call, no login. Every template targets the pinned global-catalog
asset simulo/robot/cartpole:v1, matching the current shipped scaffolds.
simulo create mybot --type training # defaultsimulo create mybot --type inferencesimulo create mybot --type scenarioPicking one
Section titled “Picking one”- Training a policy for a task you can describe with an observation, a reward, and a termination condition → training.
- You additionally want to score the trained policy over multiple rounds and
produce a recorded rollout (an MCAP flight recording of the policy’s
observations, actions, and rewards) → inference. For a rollout MCAP
with an embedded video channel on top, see the Cartpole
Eval example, which adds a camera and
include_video=True. - You’re not training anything — you want to script a deterministic simulation (a demo, a physical plausibility check, a hand-written controller) → scenario.
Where to edit
Section titled “Where to edit”Every template puts the parts you’re expected to change at the top, marked
# EDIT: in the generated code — the scene construction (build()), the
observation/reward/termination logic, and the job’s own parameters, which
are the CLI flags (no @app.entrypoint — simulo run maps flags onto the
job directly; see App & Jobs).
Everything else (the runtime image, the
volume wiring, the checkpoint callback) is boilerplate you can leave alone
until you need to change it.
Once you’ve picked a robot to build with, see Bring your own
robot for publishing your own USD or URDF
package and replacing the global cartpole ref. The API Reference documents the
submit-side classes the templates use (simulo.App,
simulo.Runtime,
simulo.Volume, the
callbacks) and the contracts they implement
(simulo.Task, simulo.Scenario,
and the component protocols), plus the GPU-runtime classes
(simulo.Robot, simulo.RLTrainer,
…) — each has its own reference page now, but the templates’ own annotated
walkthroughs are still the better first read for how those classes fit
together end to end.