stray-light/zerograv-manipulation-regular-gravity-trajectories-state-v1

Regular-Gravity (Vanilla-Matched) Manipulation Trajectories (State-Only, v1) Dataset Card

The regular-gravity control arm for stray-light/zerograv-manipulation-trajectories-state-v1. Same DROID rig, same tasks, same PPO recipe, same per-task QA filters, same hold-segment trimming, same LeRobot layout, and ~1000 episodes per task — the only intended difference is the physics: gravity is always on, whereas the zero-gravity (ZG) datasets float the object (fully, or until contact/grasp). The point is to give a distillation study a matched baseline: identical spawn distribution, identical expert-training recipe, different gravity.

Tasks, checkpoints, and dataset stats

Task	Episodes	Frames	Checkpoint	Peak eval success_once	Mean of last 10 evals
PushCube	1000	178,524	droid-pushcube-vanilla-matched-armbound02-full50M-h300/best_ckpt.pt	0.875	0.72
LiftPeg	1000	258,572	droid-liftpeg-vanilla-matched-armbound02-full50M-h425/best_ckpt.pt	1.00	0.85
PickCube	1000	449,615	droid-pickcube-vanilla-matched-armbound02-full50M/best_ckpt.pt	0.688	0.36
PickSingleYCB	1000	466,032	droid-picksingleycb-vanilla-matched-armbound02-full80M-h900-ne2048/best_ckpt.pt	0.625	0.51

Frames is the sum of each task's 1000 trimmed episode lengths (see "Hold-segment trimming"). Eval success numbers come from the training run's own 16-env in-training eval, which is noisy; every training run finished its full budget (unlike the PickSingleYCB ZG checkpoint, which came from a run still in progress). PushT and PegInsertionSide are not included in this release.

What "Vanilla-Matched" means

Env ids: PushCube-DROID-Vanilla-Matched-v1, LiftPeg-DROID-Vanilla-Matched-v1, PickCube-DROID-Vanilla-Matched-v1, PickSingleYCB-DROID-Vanilla-Matched-v1. Each is the corresponding ZG DROID env with disable_gravity=False (real gravity, no on-contact / on-grasp latch), the same containment walls, and match_3d_spawn=True — the object spawns in the same 3D box the ZG env uses (objects start floating in mid-air and fall), with the same 0.5 m/s linear / 1.0 rad/s angular initial kick; goals are unchanged from the ZG env. Isolating gravity as the only variable is why the spawn is matched rather than table-surface.

Per-task training hyperparameters (PPO, ManiSkill baseline)

Identical to the ZG v1 recipe (values read from each run's logged config): control_mode=hybrid_impedance_delta, arm_action_bound=0.2, gamma=0.95, gae_lambda=0.9, clip_coef=0.2, lr=3e-4, num_steps=50, obs_mode=state, seed=1.

param	PushCube	LiftPeg	PickCube	PickSingleYCB
max_episode_steps	300	425	900	900
ent_coef	0.0	0.0	0.0	0.002
num_envs (train)	1024	1024	1024	2048
total_timesteps	50M	50M	50M	80M

Collection QA filtering

Same per-task filters as the ZG v1 collection (each kept episode also passes the stock success flag at some point): PushCube --require-success-at-end; LiftPeg --min-stable-hold-frames 15 (+ default settle check); PickCube and PickSingleYCB --settle-vel-threshold 0 --require-grasp-at-success --min-stable-hold-frames 15. All collected with --full-horizon, 256 envs, seed 0.

Task	Episodes attempted	Raw successes	Dropped by hold/end filter	Kept (before exact-1000 cut)
PushCube	2,816	2,067 (73.4%)	903 not in success on final step	1,164
LiftPeg	3,584	3,035 (84.7%)	2,019 final success run < 15 frames (+1 never settled)	1,015
PickCube	5,120	2,061 (40.3%)	1,014 final success run < 15 frames	1,047
PickSingleYCB	4,864	2,064 (42.4%)	823 final success run < 15 frames	1,241

The first 1000 kept episodes (in collection order) are shipped.

One intentional difference: the ejection filter threshold

The collector's ejection filter drops episodes whose object speed exceeds a threshold within the first 5 steps of a reset (it catches PhysX depenetration blow-ups). The ZG collection used the default 1.0 m/s. With gravity on and a floating 3D spawn, objects legitimately free-fall: on PushCube, 95% of zero-action resets exceeded 1 m/s (100% for spawn height >= 0.2 m; max downward speed 2.98 m/s, below the 3.27 m/s free-fall bound over 5 steps at 15 Hz), versus 0% for the ZG env. At 1.0 m/s the filter therefore acts as a spawn-height filter and kept only 0.4-1.6% of episodes in a smoke test. This collection uses --ejection-vel-threshold 4.0, above free-fall plus the initial kick and far below the ~21 m/s peaks of real depenetration ejections measured earlier in this project. At 4.0 the ejection filter dropped 0 episodes, and the initial-object-height distribution matches the ZG datasets (mean 0.29-0.33 m in both arms; 69-83% of spawns above 0.2 m in both).

Consequence for users: vanilla episodes begin with the object falling from its random spawn height, so the first few frames contain fast downward motion (object speed p99 over the first 5 steps is 2.6-2.9 m/s, vs 0.75-1.0 m/s in the ZG data). This is real gravity, but a model will see it.

Hold-segment trimming

Identical to ZG v1: every episode is collected to its full horizon, then trimmed to (start of the final contiguous success run) + 15 frames (~1 s at 15 Hz), so a 15-step chunked policy still sees genuine stop-and-hold behavior. Every trimmed episode ends on a success frame.

Comparison with the ZG v1 datasets (measured on the shipped data)

Task	Attempts to keep 1000 (ZG -> regular)	Trimmed frames/episode	Mean peak object height, m	Median steps to first success
PushCube	4,352 -> 2,816	220 -> 179	0.359 -> 0.304	181 -> 153
LiftPeg	18,176 -> 3,584	291 -> 259	0.423 -> 0.326	240 -> 198
PickCube	6,912 -> 5,120	603 -> 450	0.418 -> 0.367	220 -> 245
PickSingleYCB	13,568 -> 4,864	527 -> 466	0.413 -> 0.360	288 -> 263

ZG trajectories reach a higher object peak on all four tasks (14-30%). Attempt counts and timings also reflect how stably each individual expert holds success, not only the gravity condition; there is one expert seed per task per arm, so treat these as descriptive rather than statistically tested.

Integrity checks performed

Run on all four exports (and on the ZG v1 exports, for reference): episode and frame counts match info.json; global/episode/frame indices are contiguous and timestamps equal frame_index/15; no NaN/Inf in state, action or reward; arm actions never exceed arm_action_bound (max |arm action| = 0.200); every episode's parquet rows equal the raw trajectory.h5 (exact match on actions and state); no duplicate episodes (same initial object state + first 50 actions); checkpoint sha256 in provenance.json matches the checkpoint file; no secret-like strings or absolute home paths in shipped text files. Additionally, each of the four regular-gravity exports loads with the lerobot library.

Key Notes

• provenance.json records git.dirty = true: this data was produced from an uncommitted working tree. The exact env/training code is shipped in each task's env_code/ folder, so the collection is reproducible from the snapshot.
• Each task folder also contains the raw trajectory.h5 / trajectory.json (full per-step env_states, untrimmed), for independent reconstruction.
• Checkpoints are not published here; they live in the training runs' runs/ directories and the Weights & Biases project Post_Recovery_Experiments.
• observation.state layout matches the ZG v1 exports task-for-task (PushCube 43-dim, LiftPeg 40, PickCube 50, PickSingleYCB 53); this pipeline never transforms the state vector, only which frames are kept.