Instructions to use torchrl/microduck-skills with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- Microduck
How to use torchrl/microduck-skills with Microduck:
# Replace SLOT with the slot specified in the model card (walk, stand, sitstand, ground_pick, kick_left, kick_right, roulade). sudo robotctl policy load SLOT torchrl/microduck-skills
- Notebooks
- Google Colab
- Kaggle
Commit History
-
Curate canonical MicroDuck artifacts
Verified signatureKeep one validated nine-skill prior and one checkpoint/video pair per game. Preserve the prior hyperparameters, evaluation, lineage, and hashes; add focused READMEs and replace duplicate tag videos with one held-out montage.d676623
-
Publish initial MicroDuck tag pilot and held-out evaluation
Verified signaturee9fd839
-
Document pinned football results and portable checkpoint loading
Verified signaturec69cf0d -
Add evaluated nine-skill football experiment and paired video
Verified signature0bb7f4b -
Clarify the new nine-skill football training and preserve selector pairings
Verified signature1366e6b -
Pin nine-skill artifact revision and document compatibility
Verified signature69892bc -
Publish validated nine-skill MicroDuck prior and labeled demonstration
Verified signature01ebcef
-
Seven-skill walker: turning in place and a level head
Verified signatureReplace the five-skill football walker with the seven-skill walker trained on 2026-09-12 (12,009,472 transitions, 16 envs, football-iter2 commit c37fd1814): standing, forward 0.25 m/s, backward -0.2 m/s, sidesteps +-0.2 m/s, turning +-1.0 rad/s, every task paying the head_level term. walker.ckpt (sha256 cbb5023d70bac278b3d066914289698c4ca53c6c8502ee39f5285fe51cf62c31), skills.json (56-episode evaluation), videos/skills.mp4 (3x3 grid of seven deterministic 10 s rollouts), training.json and the model card, which records the measured football-scene numbers and the backward head-drop limitation. navigation.ckpt, navigation.json and videos/navigation.mp4 are unchanged.8c31e26 -
Publish the five-skill football walker and its skills video
Verified signaturewalker.ckpt: 10,002,432-transition PPO walker with the football task library (standing, forward 0.25 m/s, backward -0.2 m/s, sidestep left and right 0.2 m/s, no jump), sha256 855380a5692fc196768cfa9bc4d813f7995f0c814fa4646233e07e34248a3f76. skills.json: its 40-episode per-skill evaluation. videos/skills.mp4: deterministic five-tile playback. training.json and README.md describe the new walker and the training recipe. navigation.ckpt, navigation.json and videos/navigation.mp4 are unchanged and remain paired with the previous six-skill walker (revision 4191d7d25c4fd58a5c6e6395fcf8217459fdd073).43ffe3b
-
Show the trained navigation policy with an external overview camera
Verified signature5824a85 -
Pin the evaluated navigation checkpoint in the loading example
Verified signature4b462a3 -
Publish evaluated MicroDuck navigation policy and final video
Verified signature6330ae6 -
Publish the evaluated 10M-transition MicroDuck skill policy
Verified signature4191d7d -
Document ongoing MicroDuck training and planned checkpoint publication
Verified signaturead57015 -
initial commit
Verified signature4479f7c