CoolFace
Datasetpublic

alfayoung/robomme_1cuben_allcases4

robomme_1cuben_allcases4 (VideoUnmaskSwap1CubeN) A single red cube is hidden under one of three cups; the cups are shuffled a variable number of times and the robot must pick the cup now hiding the cube. Prompt is color-free: "watch the video carefully, then pick up the container hiding the cube". Exhaustive depth-4 enumeration. Unlike the sibling robomme_1cuben_allcases (which samples 120 cases of length 0..3 three times each), this set enumerates every discrete case up to 4… See the full description on the dataset page: https://huggingface.co/datasets/alfayoung/robomme_1cuben_allcases4.

sourceHugging Faceapache-2.0updated 1mo agoView on Hugging Face
0likes194downloads
Dataset Card

robomme1cubenallcases4 (VideoUnmaskSwap1CubeN)

A single red cube is hidden under one of three cups; the cups are shuffled a variable number of times and the robot must pick the cup now hiding the cube. Prompt is color-free: "watch the video carefully, then pick up the container hiding the cube".

Exhaustive depth-4 enumeration. Unlike the sibling `robomme_1cuben_allcases` (which samples 120 cases of length 0..3 three times each), this set enumerates every discrete case up to 4 swaps exactly once:

  • —object init position: which of the 3 cups hides the cube → 3
  • —swap trajectory: ordered sequence of swap-pairs, length 0..4, all sequences with repeats allowed (so instant-undo trajectories are included) → Σ₍ₖ₌₀..₄₎ 3ᵏ = 121

→ 3 × 121 = 363 unique cases = 363 episodes (139,251 frames). Swap-count distribution: {0: 3, 1: 9, 2: 27, 3: 81, 4: 243} — 67% of episodes are 4-swap, the longest trajectories.

Fixed cups. Every episode is generated with the cups frozen at their exact nominal lanes (no per-episode row tilt or jitter). Combined with one realization per case, the only axes that vary across the 363 episodes are (init position, swap sequence, recovery mode) — there is no continuous cup-pose nuisance at all.

Uniform recovery. Recovery mode is drawn uniformly over {none, z, xy} (122 / 121 / 120 episodes), so ~2/3 of episodes contain a deliberate failed grasp plus recovery. Note this differs from the 0.70 / 0.15 / 0.15 weighting used by the sibling robomme_1cuben_allcases, so recovery statistics are not comparable across the two.

Verification. Demo-segment length is a deterministic function of swap count (114 / 114 / 168 / 216 / 264 steps for 0 / 1 / 2 / 3 / 4 swaps — a clean +48-step swap window), which was used to confirm that the depth-4 cases execute four real swaps and that episode ordering is preserved end to end. All 363 cases are distinct and present. Seeds 900000-936200, disjoint from the official/train splits of the sibling datasets.

Standard LeRobot v2.1 layout (panda, 10 fps). Precomputed VAE latents are not included.