GeorgiaTech/ec-latent-playkit

EC latent playkit

Self-contained inputs for the latent-playground notebooks in the Embodied-Control repository (branch feat/latent-perturbation-playground):

• notebooks/z256_latent_perturbation.ipynb — continuous 256-dim latent (bundles/rollout24_gamma097_3500m)
• notebooks/fsq64_latent_perturbation.ipynb — quantized 64-dim FSQ latent (bundles/fsq64_sonic_4500m)

bundles/<name>/  exported policy bundle: TorchScript tracker + skill encoder,
                 observation/action contracts, normalizer, golden trace,
                 provenance (training-checkpoint SHA)
reference/       reference-array tree (root_qpos_v1) with the motions to encode
model/           G1 MJCF plus its meshes, for the MuJoCo plant and the renderer
playkit.json     what this kit was built from + per-file sha256

Motions

reference/root_qpos_v1 carries the 30-motion bones_seed_language30_compositionality_v1 set (14,423 frames at 50 Hz): locomotion, manipulation, and idle/gesture clips. Two of them are known to be tracker-limited — the oracle latent itself falls on 4 of 5 evaluation episodes in the training simulator, so a fall there says nothing about your perturbation:

• panic_run_away_180_R_001_A423
• walk_big_dog_ff_225_stop_R_001_A492

Want different motions?

The full processed motion catalog is public — browse it and pick by name:

• GeorgiaTech/g1_bones_seed_sonic_129k_50hz — all 129,785 SONIC BONES-SEED clips retargeted to the G1 (sharded NPZ + g1_bones_seed_sonic_full_manifest.json with every clip name, and a language-description JSON).
• GeorgiaTech/g1_bones_seed_100_50hz — a curated 100-clip subset.

The playground needs motions converted into a reference-array tree, which is done in the training repository (python -m imitation_experiments.data.build_reference_arrays); send a list of clip names and a rebuilt kit revision can include them.

Use it

This dataset is public — no account, token, or org membership is needed. The notebooks download the kit automatically, pinned to an exact revision, so nothing here needs to be fetched by hand. From the Embodied-Control checkout:

./scripts/setup_latent_lab.sh

To fetch manually instead, either use huggingface_hub:

from huggingface_hub import snapshot_download
snapshot_download("GeorgiaTech/ec-latent-playkit", repo_type="dataset",
                  revision="<pinned sha from the notebook>",
                  local_dir="assets/latent_playkit")

or plain HTTPS, one file at a time (playkit.json lists every path and its sha256):

curl -sL "https://huggingface.co/datasets/GeorgiaTech/ec-latent-playkit/resolve/<revision>/<path>" -o <path>

Provenance

Every training checkpoint SHA is in bundles/<name>/manifest.json (source.checkpoint_sha256, source.skill_checkpoint_sha256), and playkit.json carries a sha256 for every file in the kit. The bundles replay a golden trace at load time (verify_bundle), so a corrupted or mismatched download fails loudly.

Nothing here is a paper metric: the plant is MuJoCo, not the training simulator, and the notebooks run single deterministic episodes.