license: apache-2.0
task_categories:
- robotics
tags:
- LeRobot
- robomme
- shell-game
robomme_1cuben_fixedcup_split (VideoUnmaskSwap1CubeN — fixed cups, disjoint split)
A single red cube is hidden under one of three cups; the cups are shuffled a variable number of times (0..3) 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".
Derived from robomme_1cuben_allcases,
with two changes for a clean generalization study:
1. Fixed cup locations. Cup-pose perturbation is removed (no_cup_perturb): the three
cups sit at their exact nominal lanes (x=0, y ∈ {-0.25, 0, +0.25}) in every episode, so the
only variation across episodes is the discrete swap trajectory — the layout confound is
removed.
2. Train/test disjoint in state transitions. The 120 unique discrete cases (3 hide
positions × every swap-pair sequence of length 0..3) are split 96 train / 24 test by a
seeded shuffle (seed 0). Every swap-pair, sequence length, and hide position still appears in
training; only the specific full (hide, swap-sequence) combinations are held out —
a compositional-novelty split. Each kept case is drawn 3× from recovery mode
{none, z, xy} (weights 0.7 / 0.15 / 0.15; ~30% contain a deliberate failed grasp + recovery).
This repo is the TRAIN split: 288 episodes (89119 frames). The held-out 72-episode test
split (the 24 disjoint cases) is listed in
record_dataset_VideoUnmaskSwap1CubeN_test_metadata.json and shipped in the RoboMME benchmark
(env_metadata/test, task index 17); it is re-simulated at eval time with the same fixed cups.
The three samples of a case share an identical initial scene but are three distinct demonstrations — the oracle planner's path (and, when present, the recovery grasp) differ.
Format
Standard LeRobot v2.1 layout (Franka Panda, 10 fps). Action = 8-dim absolute joint
position (pd_joint_pos: 7 arm joint angles in radians + 1 gripper open/close). state is
the matching 8-dim joint observation. Precomputed VAE latents are not included (they are
model-specific; regenerate with the RoboMME latent-precompute step).
Split metadata (at repo root)
record_dataset_VideoUnmaskSwap1CubeN_train_metadata.json— the 288 train records (case,seed,hide_idx,swap_pairs,n_swaps,recovery_mode,no_cup_perturb).record_dataset_VideoUnmaskSwap1CubeN_test_metadata.json— the 72 held-out test records.