Skip to content
TrainingSource reviewed

MicroDuck training to deployment: a complete reading path

Follow preparation, local training, export, simulation replay and deployment checks in order. A successful export or upload is not evidence that the robot can execute the policy.

Training to deployment

  1. MicroDuck RL Repository: Training Workflow & Export

    The official MicroDuck RL repository is the source for training setup, cloud jobs and policy workflows. Start with its current quick start, then record the revision and compute environment used for an experiment. Our site provides preparation notes and does not run training jobs.

  2. MicroDuck Local Training: Setup, Checks & First Run

    Start with a separate checkout of the reviewed official revision, sync its locked dependencies with Python 3.12, and check CUDA from that environment. Run a small installation check before committing compute to a full training session.

  3. MicroDuck ONNX Export: Checkpoints, Replay & Validation

    Export with the official MicroDuck exporter so observation normalization stays inside the ONNX model. Current policies use 61 inputs and 14 actions; local replay needs --new-cmd-obs for that layout. Validate and rehearse the exported file before sharing it.

  4. MicroDuck Simulator: First Walk and Public Move Loading

    Open the official simulator and choose Waddle In to try the built-in walk. A ?move=org/repo URL can load a public move; the Playing label identifies what is loaded, not whether it is driving. Its X button restores the official slot. We checked the public sprint reference load and unload on October 12.

  5. MicroDuck Policies: Update v8, Inspect Overrides & Reset

    The official MicroDuck policy set added velstand v8 on October 12, but a daemon upgrade does not prove your robot runs it. Compare robotctl policy list and policy check, then update the official set if appropriate. A manually loaded walk override remains in place until you intentionally reset that slot.

Read the bundle before treating ONNX as training input

The RL changes are merged at bae39f59. The public training repository at 64d28ce0 lists nine bundles; this is newly shared provenance, not nine new measured robot results. The reviewed policy set remains v8. An ONNX actor export does not supply the critic, optimizer and curriculum state needed for normal resume.

The velstand example is iteration 2749 from an October 12 run, linked to v8 velstand.onnx, training commit 37507767 and two teacher bundles. The teachers record on_remote:false, although both code commits were directly readable on GitHub in this review; that flag alone does not mean missing source. Roulade has no recorded code commit. Velstand’s uncommitted.diff contains clean Git status and no code patch: inspect it rather than applying every file blindly.

  1. Match task_id, checkpoint_file and checkpoint_sha256; distinguish checkpoint bytes from the ONNX digest.
  2. Inspect code.commit, on_remote, configuration and the actual content of any diff file. A recorded training commit and a newer resolver commit serve different purposes.
  3. Check every teacher bundle and needed source revision. Public weights alone do not supply missing code, access or environment details.

Choose resume, warm start or sharing deliberately

The new resolver accepts local .pt files or hf://owner/repo@revision/path. Same-task --init-checkpoint resumes iteration and curriculum counters; MICRODUCK_WARM_START=1 restarts curricula for another task while retaining loaded model state. Avoid mixing the new entry with the old resume flags. The recorded training checkout may predate this CLI: verify its supported entry before copying the newer example.

Sharing happens only with share-training or publish --share-training, and newly created repositories default to private unless --public is chosen. Review provenance and attached diffs before publishing. This site read code, nine metadata records and small reference files; it did not download or load checkpoints, train, reproduce a gait or validate HD1910 compatibility.

Choose by the task you want to finish.

Choose a starting environment
Your goalStart hereWhat to check
Try walking and the controlsBrowser simulatorWait for the scene to load. No local training setup is needed.
Train on Linux with an NVIDIA GPULocal training preparationPython 3.12, the locked dependencies, driver access and CUDA visible inside the project environment.
Train from Windows with an NVIDIA GPUReview WSL 2 as a candidate Linux environmentFollow NVIDIA’s WSL guidance, then run the project checks inside WSL. CUDA support alone is not proof of full MicroDuck compatibility.
Use a Mac or a computer without CUDABrowser exploration; cloud compute for the official training routeA working browser or a Python install does not establish support for the official CUDA training backend.
Use DGX Spark, GB10 or Jetson hardwareRead the dependency sources for the exact boardARM is not one interchangeable environment. Jetson Thor has a separate open dependency report.
Replay your own exported policyExport and local replay guideMatch the input layout, model and control settings. CPU inference does not make dependency installation platform-independent.

The environment requirements come from MicroDuck RL quick start and task registry ↗ and Python, dependency pins and ARM package sources ↗. Windows guidance is a candidate route based on NVIDIA: CUDA on WSL 2 ↗, not a MicroDuck test performed by this site. For Thor, read Jetson Thor report #38 (community report, open at review) ↗ before applying advice intended for GB10.

What a GPU check can tell you.

A system utility seeing an NVIDIA GPU and Python being able to use it are two separate checks. Run the Python probe in the local setup guide from the same project environment that will run training. Record the GPU model, available memory, driver, Python version and revision before comparing your result to somebody else’s.

The official 4,096-environment example describes parallel simulated environments, not a universal memory requirement. Start smaller to check the installation. Do not treat a short successful run as a trained walking policy or use an estimated training time as a guarantee for your machine.

Plan a cloud job before starting one.

The official wrapper adds --hf-jobs to a training command. It also exposes the hardware flavor, namespace, timeout and a dry-run option. Read MicroDuck cloud job submission guide ↗ and check Hugging Face Jobs: current pricing and billing ↗ for current eligibility and compute charges.

  1. Choose the personal or organization namespace that should own the run and its bill.
  2. Choose the GPU and an explicit time limit.
  3. Check where the checkpoints will be stored.
  4. Preview the submission, then decide whether to start the paid job.
  5. Keep the job URL so you can check completion and cancel it if needed.

Keep an experiment small enough to explain.

Begin with an unchanged official walking task. Save the initial configuration and a playback observation. For a second run, change one setting and explain what you expected it to affect. This gives you a useful comparison even when the second result is worse.

Continue with your result.

A separate CPU route for Mac experiments.

Microduck Lab by jonathanhawkins ↗ is an independent project updated September 22, 2026. Its author provides an Apple Silicon CPU training harness and a setup script that pins upstream model and policy revisions, prepares Python and viewer dependencies, and runs contract checks. Start with a fresh checkout and read its requirements before running the script. This is useful for local reward-design experiments; its reported results and claims of compatibility are the author's, not a validation by this site or the official CUDA stack.

Keep the source revision, training environment, exported ONNX and evaluation results beside any comparison. The official daemon simulation answers a different question: how the robot software behaves against a simulated body.

CONTINUE THE CONVERSATION

What would you like to understand next?

Keep the source, your environment and your observations together.

Before you continue

Can a deployed ONNX file resume training?

An ONNX actor is not a complete training checkpoint. Check the optimizer, critic, curriculum, code revision and required teacher records.

Applies to: Pollen Robotics MicroDuck. Keep the model, image, software and driver versions together. The dated sections below retain their original verification limits.

Source checked:

Sources and review (16)
  1. github.com/pollen-robotics/microduck_rl/blob/2b25a48b08f1f17bc38c90bb03144c81fbd9ed07/README.md
  2. github.com/pollen-robotics/microduck_rl/blob/2b25a48b08f1f17bc38c90bb03144c81fbd9ed07/pyproject.toml
  3. docs.nvidia.com/cuda/wsl-user-guide/index.html
  4. github.com/pollen-robotics/microduck_rl/issues/38
  5. github.com/pollen-robotics/microduck_rl/blob/2b25a48b08f1f17bc38c90bb03144c81fbd9ed07/scripts/hf/README.md
  6. huggingface.co/docs/hub/jobs-pricing
  7. github.com/jonathanhawkins/microduck-lab/tree/47c62e3fb5193191d6d7e1181cc61089eaa768c7
  8. github.com/pollen-robotics/microduck_rl/blob/bae39f5900b318f7e87fc485c30a31213cfa14eb/docs/training_checkpoints.md
  9. huggingface.co/pollen-robotics/microduck-training/blob/64d28ce092253c509b51031ec7ccb53790a00d0d/README.md
  10. huggingface.co/pollen-robotics/microduck-training/blob/64d28ce092253c509b51031ec7ccb53790a00d0d/bundles/velstand/0045i2si_2749/bundle.json
  11. github.com/pollen-robotics/microduck_rl/blob/bae39f5900b318f7e87fc485c30a31213cfa14eb/src/mjlab_microduck/checkpoints.py
  12. github.com/pollen-robotics/microduck_rl/blob/bae39f5900b318f7e87fc485c30a31213cfa14eb/src/mjlab_microduck/publish/training.py
  13. github.com/pollen-robotics/microduck_rl/blob/bae39f5900b318f7e87fc485c30a31213cfa14eb/src/mjlab_microduck/train_hook.py
  14. huggingface.co/pollen-robotics/microduck-policies/blob/665ae4eb20bfe13bca1ef4ea6d4bb6b4a096e448/manifest.json
  15. github.com/pollen-robotics/microduck_rl/commit/85105f4b135e548f4df791227d2eed12de0bef2c
  16. github.com/pollen-robotics/microduck_rl/commit/2f1f76bc6abb4df32ebf46510215b3d2c497c2b9

Next useful guides

MicroDuck Local Training: Setup, Checks & First Run

Prepare the official MicroDuck RL environment with a pinned revision, Python 3.12, a CUDA check and a short training run before a longer experiment.

MicroDuck ONNX Export: Checkpoints, Replay & Validation

Export a MicroDuck checkpoint with observation normalization, select the correct playback input layout and inspect a policy package before sharing.

MicroDuck Reference Policies: Sprint, Rollers & Sit/Stand

Compare four official reference policies by slot, accessories and command meaning. Inspect pinned manifests, try a matching model and reset the override.