text stringclasses 3
values |
|---|
2e1fbefbea31ba6ee6e07c3fb97eb4d2ffe6f9f8f89aa60d94e50cccaca2e044 taskdemo_l2l3.tar |
396c5bddc4a13399e5d62749314eac51c29402736b535ffdb0adca8495a551fd pools/dth_pool_1x_v6.npz |
441d54d3a05c2186cc6a1c5fedbba46a033704870fd63521060b030adce419aa pools/taskdemo_pool_v6.npz |
mimicdroid-gr1
GR1 one-shot post-training set: 579 zarr episodes (4.7 GB unpacked) plus the retrieval manifest they
are matched with. Two embodiments share the same kitchen tasks β a floating-hand teleop demonstrator
(mimicdroid) and the GR1 humanoid (gr1) β so a policy can be conditioned on a demo from one and act
with the other.
| count | |
|---|---|
| episodes | 579 (gr1 253, mimicdroid 326) |
| windows | 93,371 |
| queries | 61,686, top-6 matches each |
| tasks | TurnOnFaucet 50, CloseLeftCabinetDoor 50, PnPSinkToCabinet 81, PnPSinkToRightCounterPlate 72, dth floating-hand pool 314, taskdemo 12 |
Layout
data/<episode>.tar # one tar per zarr episode (uncompressed; zarr chunks are already compressed)
manifest/ # retrieval manifest + normalization statistics
windows.npz # ep / w0 / w1 / emb / cls / keep for every window
pairs.npz # query_idx, match_idx (N, 6), match_dist
state57_stats*.json # state normalization (percentile-hybrid variant used by the v6 recipe)
action57_stats.json, abs_action57_stats_pctl.json, cmd_delta57_stats.json
t5_embeddings.pkl # per-episode T5 text embeddings
episodes.txt # the 579 episode names, in manifest order
stats.json # the table above, machine-readable
Each episode unpacks to a zarr group:
<episode>/
images.front_1 images.center_top images.sideleft images.sideright # 4 camera views
state57 action57 abs_action57 # 57-d state / action / absolute command
meta
Use
mkdir -p egoverse && for t in data/*.tar; do tar xf "$t" -C egoverse; done
Point the training config at the unpacked root and the manifest directory:
HUM_ROOT = <...>/egoverse # dataset class resolves manifest `ep` names under this root
manifest_dir = <...>/manifest
windows.npz:ep is exactly <episode> in data/<episode>.tar, so the two halves line up with no
renaming. Absolute-command targets (abs_action57) are what the v6 post-train regresses; executing the
predictions reproduces the demo tracking dynamics.
Provenance
Episodes were generated in RoboCasa kitchens: a floating-hand pool per task and level, retargeted to
GR1, then selected into the final per-task sets. The selection provenance is kept per demo
(v6_origin) in the source HDF5s, which are not part of this upload β this repository is the
training-ready half (manifest + zarr).
Eval assets (eval/)
These are the files the GR1 one-shot closed-loop eval (gr1_local_eval.sh in
jeongeun980906/watch) needs in addition to the training data:
eval/taskdemo_l2l3.tar # 25 one-shot taskdemo zarr episodes for the L2/L3 tasks (unpack into egoverse/)
eval/pools/taskdemo_pool_v6.npz, eval/pools/dth_pool_1x_v6.npz # eval_gr1_server --pool / --play-pool
eval/sha256.txt
The 12 seen-task taskdemo_* episodes are already in data/. The L2/L3 ones were built from the TaskDemos
hdf5 of HF Rutav/MimicDroidDataset, with the same converter as the seen ones. The rebuilt seen-task
episodes match the ones in data/: the arrays are identical, the pixels differ by at most 2, and the
attributes are identical. The pools are built from these episodes. With --oneshot-pool3 the server loads
them at startup but does not use them.
tar -xf eval/taskdemo_l2l3.tar -C egoverse # next to the episodes unpacked from data/
- Downloads last month
- 53