Core Concepts

You describe a robot learning problem in Python, preview the robot inside its task, then submit training with simulo run. A GPU machine in the Simulo cloud trains it. What comes back is a policy: the network that picks your robot’s actions, saved as best and latest checkpoints. Evaluation runs one of those checkpoints through fixed episodes and measures it against the success rule you wrote before you decide to play, continue training, or export it.

task.py  ─┐
          ├─ simulo run train.py ──▶  TRAIN job_brisk-heron-4f8k2m
train.py ─┘                                   │
                                              ▼
                                   policy_brisk-heron-4f8k2m
                                     ├─ :best
                                     └─ :latest

A typical project has five Python files. task.py holds your app and your task: the scene, what the robot observes, the reward, and when an episode ends. preview.py checks the robot in the task, train.py trains a policy, eval.py measures a checkpoint against your success rule, and play.py runs a checkpoint back to watch or record it. simulo create NAME writes all five for you.

You write against top-level simulo.* names such as simulo.App, simulo.Task, and simulo.Robot. On your machine these names are lightweight and need no GPU and no PyTorch, so authoring and submitting an app is fast and works on a laptop. On the GPU machine that runs your job, the same names are the full implementations. Your code does not change between the two.

Reading order

Start with Apps and jobs, then Job lifecycle, Preview, Evaluation, and Policies and checkpoints. Together they cover the path from simulo run to a measured checkpoint you can download or use. The rest can be read in any order.

Apps and jobsOne typed job per file, the job types, and what simulo run does with your file.

Job lifecycleStatuses, failure reasons, timeouts, cancelling, and the commands that show a job's progress and outputs.

PreviewCheck a robot inside its task before training, read the findings, and inspect the recording.

EvaluationMeasure a trained checkpoint against your success rule, inspect failures, compare checkpoints, and gate a pipeline.

Policies and checkpointsWhat a training job produces, how best and latest are chosen, and how to continue from, play, or export a checkpoint.

Tasks and RL trainingsimulo.Task defines the learning problem; LearningEnv runs many copies of it; RLTrainer trains a policy on it.

Scene, Robot, Prop & WorldWhat you put in a scene, and which parts each parallel environment gets its own copy of.

Asset catalogs & version pinningPublish robots, worlds, and props once, reuse them by reference, and keep every job's inputs pinned.

Asset validationThe checks a package passes before it gets a version, including the settle and actuation probes for robots.

Volumes & AssetsWritable scratch folders on the GPU machine, and read-only catalog inputs pinned at submit.

RuntimesThe prebuilt environment every job runs in, and how to add public PyPI packages and environment variables.

System typesThe GPU tiers a job can request with system=, and what that choice does.

Callbacks and anomaly captureDebugOnAnomaly records the moments before training goes wrong, without stopping the job.

Recordings & live viewDownload MCAP recordings after a run, or watch a training job in 3D while it runs.