frame_stride int64 | phase_align int64 | seed int64 | val_per_stratum dict | episodes list |
|---|---|---|---|---|
4 | 4 | 42 | {
"114": 2,
"168": 6,
"216": 16,
"264": 49
} | [
0,
1,
2,
3,
4,
5,
6,
8,
9,
10,
11,
14,
15,
17,
18,
19,
20,
21,
23,
25,
26,
27,
28,
30,
31,
33,
35,
36,
37,
39,
40,
41,
42,
44,
45,
46,
47,
48,
49,
51,
52,
53,
54,
55,
56,
57,
58,
60,
61,
63,
64,
67,
68,
69,
7... |
robomme_1cuben_allcases4_phase4
Directly trainable LeRobot-format build of the exhaustive depth-4
VideoUnmaskSwap1CubeN dataset. A single red cube is hidden under one of three
fixed cups, the cups are shuffled zero to four times, and the robot must pick the
cup hiding the cube.
This repository includes the raw LeRobot parquet data, both precomputed camera
latents, the classifier-free empty text embedding, and deterministic train and
validation manifests. Unlike alfayoung/robomme_1cuben_allcases4, no local
latent-precomputation step is required.
Coverage
- 363 unique episodes: every
(hide position, ordered swap sequence)case once - Swap-count distribution:
{0: 3, 1: 9, 2: 27, 3: 81, 4: 243} - Fixed cup locations with no cup-pose perturbation
- Uniform recovery modes:
none,z, andxy - One color-free task prompt and a red cube
Phase-4 preprocessing
Video frames are sampled with frame_stride=4. The demo and execution phases
are then trimmed independently at the tail to sampled-frame lengths equal to
1 mod 4 before separate causal Wan VAE encoding. This gives an explicit
phase-alignment match for the Wan temporal factor of four.
The stored latents were rebuilt from the original 363 parquet episodes and all 726 files were validated for episode coverage, matching camera metadata, phase-boundary geometry, and latent dimensions.
Split
The split uses seed 42 and is stratified by exec_start_idx:
splits/1cuben4_phase4_train.json: 290 episodessplits/1cuben4_phase4_val.json: 73 episodes
The manifests are disjoint and cover all 363 episodes exactly once.
Layout
data/chunk-000/ # 363 LeRobot parquet episodes
latents/chunk-000/front_rgb/ # 363 precomputed latent files
latents/chunk-000/wrist_rgb/ # 363 precomputed latent files
meta/ # LeRobot v2.1 metadata
splits/ # train and validation manifests
empty_emb.pt # classifier-free empty text embedding
dataset_provenance.json
The repository root is the dataset root expected by LingBot-VA. After
downloading it, point dataset_path directly to the downloaded directory.
Prompt: watch the video carefully, then pick up the container hiding the cube
- Downloads last month
- 174