source string | split string | seeds list | scenes int64 | stages int64 | horizon int64 | goal string | reached dict | every int64 | subgoals_per_scene dict |
|---|---|---|---|---|---|---|---|---|---|
scripted nominal + recovery | val | [
100035,
101075
] | 100 | 4 | 1,795 | first step with k placed cubes | {
"stage1": 1,
"stage2": 1,
"stage3": 1,
"stage4": 1
} | 80 | {
"mean": 18.85,
"min": 18,
"max": 22
} |
scripted nominal + recovery | val | [
100035,
101075
] | 100 | 4 | 1,795 | first step with k placed cubes | {
"stage1": 1,
"stage2": 1,
"stage3": 1,
"stage4": 1
} | null | null |
CIS6280 stacking datasets, v2 (4-cube trajectories)
20,000 episodes of a dexterous-hand cube stack in the stacking-trajectories-v0.3
format, collected so that episodes continue through a 4-cube stack. Checkpoints
trained on this data are in
YongYong/cis6280-stack-world-models-v2.
The earlier 2-cube dataset was removed from the Hub.
Episodes
bc-noise-4cube-20k/episodes-*.tar holds 500 episodes each (uncompressed). Extract
every shard inside that folder:
cd bc-noise-4cube-20k && for f in episodes-*.tar; do tar xf "$f"; done
Each episode has observations.npz (front and side RGB, proprio, executed actions
and the noise applied to them, rewards), privileged.npz (simulator state for exact
replay), metadata.json, and preview.mp4. manifest.json lists every episode
with its split, outcome, and checksums. collect.log is the collection log.
| Teacher | privileged BC, chunk 10 / execute 5, stop after 4 placements (or 1795 steps) |
| Noise | per episode: 10% σ = 0, otherwise log-uniform on [0.005, 0.2] |
| Scenes | seeds 400000–419999 |
| Split | train 15990 / val 1991 / test 2019 (scene hash) |
| Rate | states, actions, and proprio at 10 Hz; RGB every 5th step |
Derived files
| Path | What |
|---|---|
derived/jepa-cache-4cube-20k/index.json |
Frame-aligned episode index (start, length, split, σ, cubes placed). |
derived/jepa-cache-4cube-20k/actions.f32 |
One 125-dim action block (5×25 raw actions) per cache frame, float32. |
derived/latents-smwm-c3/ |
Frozen SMWM-c3 latents, float16, 192-dim, for all 7,345,132 cache frames, with the train mean and std. |
goals/ |
Goal images for 100 val scenes: stage goals (scripted-val100, 1–4 cubes) and subgoals every 80 steps (scripted-val100-every80). These are the images used in the reported planning eval. |
The frame memmap (frames.u8, 7,345,132 × 128 × 256 × 3, about 673 GB) is not in
this repo. Rebuild it from the episodes, which also rewrites index.json and
actions.f32:
python scripts/build_stack_jepa_cache.py --data bc-noise-4cube-20k --out cache-4cube-20k
latents-smwm-c3 is aligned with the uploaded index.json, so policy training can
use the uploaded latents and actions without rebuilding frames. Rebuilding the cache
reproduces the same frame order because it follows manifest.json.
- Downloads last month
- 42