MicroDuck skills for TorchRL

One recurrent task-conditioned policy covering seven skills (stand, four directions, turn each way) through a task embedding and a 128-unit GRU, trained on 12 million transitions across 16 native CPU environments, with all 56 evaluation episodes surviving the full horizon. Published by the TorchRL project, which also states its known limits: the head still drops about 63 degrees walking backward, and turns reach roughly two thirds of the commanded rate. Sim-only.

Project Overview

This repository exposes one canonical nine-skill locomotion prior and one canonical export for each published game. Each artifact has a stable path, one inference checkpoint, one video, and the metadata needed to understand how it was produced.

Artifact Checkpoint Video Status --- --- --- --- Nine-skill locomotion prior prior/checkpoint.ckpt prior/video.mp4 Validated Football selector games/football/checkpoint.ckpt games/football/video.mp4 Experimental Tag selector games/tag/checkpoint.ckpt games/tag/video.mp4 Unsolved pilot

The game selectors depend on the canonical locomotion prior; they do not contain or replace its weights. Read the artifact README before loading a checkpoint:

Task

Field	Value
Source	HUGGING FACE MODEL SOURCE
Runtime contract	61 observations / 14 actions / 50 Hz
Verification	Pending independent Duckio review
Artifact	Not indexed

Source Coverage

The upstream README includes: Nine-skill prior, Versioning.

Intended Use

Use this entry to assess the project and locate its canonical implementation. Read the complete README, inspect the manifest and artifacts, and reproduce the source's evaluation before assigning simulation or hardware verification.