Step 4: Validate the Generated Dataset

Structural Validation

Check the generated HDF5 for structural problems and summarize its contents:

GR-1G1

python scripts/validate_dataset.py ./datasets/generated_dataset_dexmimicgen_gr1.hdf5

The validator prints one summary row per file (episode count, the env id recorded in the file’s metadata, and the simulation args) followed by per-file issues. Every episode is checked for the required fields (actions, initial_state, obs).

For unreadable, invalid, or truncated files, see Invalid or Truncated HDF5.

Visual Validation

Replay generated episodes to inspect them visually using Isaac Lab’s replay tool. CPU simulation matches how the humanoid demonstrations were recorded:

GR-1G1

python submodules/IsaacLab-Arena/submodules/IsaacLab/scripts/tools/replay_demos.py \
   --task Isaac-PickPlace-GR1T2-Abs-v0 \
   --viz kit \
   --device cpu \
   --dataset_file ./datasets/generated_dataset_dexmimicgen_gr1.hdf5

A good generated demonstration looks like a plausible human one — a clean, firm grasp and a stable place, with both arms moving coherently. Common artifacts worth watching for are jerky transitions at subtask boundaries (interpolation too short — raise num_interpolation_steps) and grasps that only just succeed (consider tightening subtask_term_offset_range or reducing action_noise in the task descriptor).

GR-1G1

Note

Isaac Lab replay is PhysX non-deterministic. Isaac Lab PhysX is not deterministically reproducible across an env.reset, so replaying an episode’s recorded actions from its saved initial state can diverge from the original. Some episodes may fail to reproduce success during replay even though every episode in the dataset was a successful demonstration at generation time. Each episode satisfied the environment’s success condition when it was generated.