bimanual-pick-and-place-mimic-seed-50ep

This is the Phase-0 seed of an ablation study on the SO-101 bimanual pick-and-place task (Monomyth-Inc). It re-packages the real-teleop dataset Monomyth-Inc/bimanual-pick-and-place-gemini-50ep into the IsaacLab-Mimic HDF5 schema, ready as a source dataset for isaaclab_mimic.datagen.generate_dataset.py.

Pipeline overview

real-50ep (LeRobot v3.0)             ← Phase 0: original real teleop
        │
        ▼  adapter/lerobot_to_isaaclab_mimic_h5.py
seed_50ep.hdf5  (this dataset)       ← Phase 1A: Mimic-format seed
        │
        ▼  isaaclab_mimic generate_dataset.py
mimic-augmented dataset              ← Phase 1B: sim-augmented
        │
        ▼  Cosmos-Transfer (NIM API)
cosmos-augmented dataset             ← Phase 2: photoreal-augmented

Reference articles (ABEJA Tech Blog 2025):

• https://tech-blog.abeja.asia/entry/isaac-sim-lab-synthetic-dataset-202507
• https://tech-blog.abeja.asia/entry/nvidia-omniverse-cosmos-202508

File layout

seed.hdf5                          # 30 GB IsaacLab-Mimic HDF5 (50 episodes)
previews/demo_{0,6,12,19,25,31,39,49}.mp4   # 13-15 MB each, stride-3 mp4
README.md                          # this card

HDF5 schema

Each demo group data/demo_<i> contains:

Path	Shape	Dtype	Notes
actions	(T, 12)	float32	LeRobot SO-101 bimanual joint position commands (deg + normalized gripper)
obs/joint_pos	(T, 12)	float32	Same convention as actions
obs/top, obs/left_wrist, obs/right_wrist	(T, 240, 320, 3)	uint8	Decoded from AV1 mp4, letterboxed
obs/datagen_info/eef_pose	(T, 2, 4, 4)	float32	DH-FK on observed joints; index 0 = left, 1 = right
obs/datagen_info/target_eef_pose	(T, 2, 4, 4)	float32	DH-FK on commanded joints
obs/datagen_info/object_pose/{object_left,object_right}	(T, 4, 4)	float32	Approximated: anchored to EEF while gripper closed (real seeds lack scene-side object tracking)
obs/datagen_info/subtask_term_signals/{left_grasp_done,left_place_done,right_grasp_done,right_place_done}	(T,)	uint8	Latched edges from gripper open/close transitions; per-episode dynamic threshold

Per-demo attrs: num_samples (= T), success (= True).

Caveats — why this is a seed and not a final training set

• Object-pose approximation. Real-teleop streams do not contain ground-truth object poses. We anchor the object frame to the gripper while the gripper is closed and to a fixed table location otherwise. This is sufficient for IsaacLab-Mimic's nearest-neighbour selection but does not reflect true scene geometry.
• No initial_state. The HDF5 omits the Isaac Sim full-state dump that annotate_demos.py --auto uses to replay episodes; this dataset is pre- annotated by the offline adapter, so the replay step can be skipped.
• DH-FK approximation. End-effector poses use a hand-rolled SO-101 DH table (see adapter/lerobot_to_isaaclab_mimic_h5.py); this is accurate to ~5 mm but not the same precision as the official URDF FK.

Reproducing locally

git clone https://huggingface.co/Monomyth-Inc/bimanual-pick-and-place-gemini-50ep
python adapter/lerobot_to_isaaclab_mimic_h5.py \
    --local-root bimanual-pick-and-place-gemini-50ep \
    --output seed_50ep.hdf5 --video-size 240,320

Then to use as a Mimic source dataset:

PYTHONPATH=/path/to/validation-mimic-aug:$PYTHONPATH \
./isaaclab.sh -p scripts/imitation_learning/isaaclab_mimic/generate_dataset.py \
  --task Isaac-PickPlace-SO101-Bimanual-Mimic-v0 \
  --device cuda --headless --num_envs 1 \
  --generation_num_trials 200 \
  --input_file seed_50ep.hdf5 \
  --output_file mimic_generated.hdf5

Provenance

Source	Identifier
Real teleop dataset	Monomyth-Inc/bimanual-pick-and-place-gemini-50ep
Adapter	validation-mimic-aug/adapter/lerobot_to_isaaclab_mimic_h5.py
Visualizer	validation-mimic-aug/viz/visualize_h5_episode.py
SO-101 Mimic env (registered externally)	validation-mimic-aug/sim_envs/so101_pickplace_mimic_env.py
Robot	SO-101 bimanual (bi_so_follower), 6 DOF/arm × 2
Original FPS / cameras	30 FPS, 3 cameras (top + L/R wrist)
Episodes / frames	50 / 129 136
Conversion runtime	~37 minutes on DGX Spark (single-process PyAV decode + h5py gzip)