srlwam

Sim-to-real data for a Franka Panda two-cube stacking task: red cube on green cube, with dataset-specific cube sizes. Simulation data generated in Isaac Lab with Mimic; real data captured on the physical cell.

Stack

stack/sim/stack_v1_largercubes — 2,000 v1 demonstrations with 3–7 cm cubes

V1 visual randomization: 400 nominal, 1,000 mild and 600 stronger-shadow episodes. Equal cube sizes sampled per shard, two RGB-D cameras, zero rotation commands and size-aware physical audits.

Dataset and documentation · File hashes · Preview

stack/sim/stack_v1 — Stack Cube v1 dataset (2,000 demos)

Calibrated against the real setup: fixed 30.48 mm cubes, zero rotation commands, two 256×256 RGB-D cameras at 10 Hz. 400 nominal, 1,000 mild and 600 stronger-shadow demonstrations; 100 shards with 20 demos each. Completed 2026-09-24, published 2026-09-26.

Dataset and documentation · File hashes · Train/validation splits · 100-demo preview

Legacy preview grids from the previous stack_v0 revision (before the 2026-09-24 replacement). See the current dataset folder for updated spawn-position comparisons.

Table camera	Wrist camera

stack/sim/stack_v0 — 2,000 demonstrations (replaced 2026-09-24)

Real-aligned cube spawning: X 0.48–0.59 m, Y −0.10–0.10 m; cube sizes 30.48 mm ±8%. All 2,000 final stacks physically verified. Twenty RGB-D HDF5 batches, 100 demonstrations each, including full state and actions. Full recorded trajectories, without first-frame trimming. See README and SHA-256 manifest.

stack/sim/stack_v0_largercubes — 2,000 demonstrations (added 2026-09-24)

Same wider spawn rectangle, cube sides randomized from 3–7 cm per batch, size-aware nonoverlap spacing and placement heights. All 2,000 final stacks physically verified; 3, 5 and 7 cm pilots passed. Twenty RGB-D HDF5 batches, 100 demonstrations each. See README and SHA-256 manifest.

stack/sim/stack_v0_paired — 4200 trajectories, 200 files, 96.75 GB

The same task, restructured for paired comparison instead of scale: 200 base demonstrations, each replayed 20 more times under domain randomization from the identical initial state and action sequence (verified bit-identical across all 200). One file per base demo, episode_N.hdf5 (~455 MB), holding the base plus its 20 variants. No depth channel — see the README inside the folder for exactly what varies between variants and what's held fixed.

stack/real/stack_v0 — 20 episodes, 20 files, 6.0 GB

Real-robot captures on the physical cell, one file per episode, 111-336 frames each. RGB-D from both cameras (raw stream plus a centre crop), joint and end-effector state. The recorded actions carry no rotation.