HumanoidMimicGen G1 Drill Pnp: train / success
Pick a drill from one table and place it on a second table. (HMG environment LMDrillPnP90.)
The success episodes of the train split: 505 episodes, 584,042 frames
at 50 Hz, one 480x640 ego-view RGB video each, with robot state, actions and per-frame
task-success labels. LeRobot v3.0 format.
How this split was made
Derived from jellyho/hmg-g1-drill-pnp-reward by a group-level 90/10 train/test split.
In the parent dataset each source demo contributes a clean replay plus its perturbed
siblings -- same seed, same grasp_step, same initial state, differing only in the
perturbation applied at t0. Splitting those across train and test would put a failure
episode in test whose matched success twin was trained on, so the split is drawn over
source demos, never over individual episodes: every episode sharing a
(source_shard, source_episode) key lands on the same side.
The draw is stratified on each group's success/failure makeup, so both sides keep the parent's class balance and every perturbation type appears in both. Seed 42. Train/test source-demo overlap is 0.
meta/episode_labels.jsonl carries orig_episode_index so any episode traces back to
the parent dataset, alongside its perturbation, first_success_step, grasp_step
and release_step.
Features
observation.state and action.joint are the 43 G1 joint angles, named in info.json:
| index | group |
|---|---|
| 0-11 | legs (6 per leg) |
| 12-14 | waist yaw/roll/pitch |
| 15-21 | left arm |
| 22-28 | left hand (Dex3-1: index 2, middle 2, thumb 3) |
| 29-35 | right arm |
| 36-42 | right hand |
Neither carries object state; the full MuJoCo sim state is separate, in
observation.sim.mujoco_state.
Companion datasets
chomeed/hmg-g1-drill-pnp-reward-train-success<- this onechomeed/hmg-g1-drill-pnp-reward-train-failurechomeed/hmg-g1-drill-pnp-reward-test-successchomeed/hmg-g1-drill-pnp-reward-test-failure
Notes
Success is the environment's own predicate (env.is_success()["task"]), evaluated every
step. Perturbed episodes that still succeeded count as successes, so "physics after a
branch point" is not a shortcut for failure -- the success side contains perturbed
episodes as well as clean ones.
- Downloads last month
- 324