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

Zero-Gravity Manipulation Trajectories (State-Only, v1) Dataset Card

Fourth collection round on the DROID rig, superseding stray-light/zerograv-manipulation-trajectories-state-v0 — same general recipe (state-only, ~1000 episodes/task, QA-filtered), but with a recorded-action bug fixed, hold-segment trimming applied, corrected QA filtering per task, and one new task (PickSingleYCB). Full per-task hyperparameters below, sourced directly from each training run's own logged hyperparameters/text_summary (not reconstructed from memory).

Known issues (found 2026-09-28, data not re-collected): the PullCube episodes were collected under a simulation bug (cube dynamics at 1/5 speed; confinement walls not restored between episodes), and objects in the five tasks collected before 2026-09-30 use restitution 0.90 instead of ManiSkill's stock 0.00. PlaceSphere (added 2026-09-30) and RollBall (added 2026-10-01) and PushT (added 2026-10-02) use stock physics and a first-success trim; RollBall also has a post-contact stillness filter. See their sections. See Known issues before using PullCube, or before comparing it with its regular-gravity twin.

Tasks, checkpoints, and dataset stats

Task	Episodes	Frames	Checkpoint	Peak eval success_once
PushCube	1000	220,307	droid-pushcube-armbound02-full50M-h300/best_ckpt.pt (training run · model)	0.75
LiftPeg	1000	291,455	droid-liftpeg-armbound02-full50M-h425/best_ckpt.pt (training run · model)	0.688
PickCube	1000	603,005	droid-repro-ourbranch-armbound02-full50M/best_ckpt.pt (training run · model)	1.00 (noisy 16-env eval)
PickSingleYCB	1000	526,846	droid-picksingleycb-armbound02-full80M-h900-ne2048/best_ckpt.pt (training run · model)	~0.5-0.875 (still-training run; see note below)
PullCube	1000	187,443	pullcube-droid-zerograv-oncontact-confined-armbound02-h300-ent001-s1-50M/best_ckpt.pt (training run · model)	0.75 (noisy 16-env eval, last-10-eval mean 0.34)
PlaceSphere	1000	376,672	placesphere-droid-zerograv-ongrasp-wallrelease-v2-armbound02-h900-s1-50M/best_ckpt.pt (training run · model)	1.00 (16-env; last-10 mean 0.86; 256-env offline 0.977)
RollBall	1000	274,621	rollball-droid-zerograv-oncontact-sbr-corridor-still-w3-armbound02-h400-g099-smooth1-ne4096-ns16-s1-50M/final_ckpt.pt (training run · model)	0.375 (16-env; final checkpoint 0.219 in a 256-env offline eval)
PushT	1000	190,015	pusht-droid-zerograv-oncontact-v3-s2-100M/best_ckpt.pt (training run · model)	1.00 (16-env; last-10 mean 0.85; 256-env offline 0.836)

PullCube was added in a later collection round (2026-09-24), after the original four tasks — same recipe and QA/trimming conventions throughout, called out individually below where its parameters differ.

Frames is the sum of each task's 1000 individual trimmed episode lengths (see "Hold-segment trimming" below) — not a fixed count per episode.

Per-task training hyperparameters (PPO, ManiSkill baseline)

All use control_mode=hybrid_impedance_delta, arm_action_bound=0.2, gamma=0.95 (RollBall and PushT: 0.99, see their sections), gae_lambda=0.9, clip_coef=0.2, lr=3e-4, obs_mode=state — the ManiSkill PPO baseline otherwise unmodified. Per-task horizon and entropy:

param	PushCube	LiftPeg	PickCube	PickSingleYCB	PullCube	PlaceSphere	RollBall	PushT
env_id	PushCube-DROID-ZeroGrav-OnContact-v1	LiftPeg-DROID-ZeroGrav-OnGrasp-WallFix-v1	PickCube-DROID-ZeroGrav-v1	PickSingleYCB-DROID-ZeroGrav-OnGrasp-v1	PullCube-DROID-ZeroGrav-OnContact-Confined-v1	PlaceSphere-DROID-ZeroGrav-OnGrasp-WallRelease-v2	RollBall-DROID-ZeroGrav-OnContact-SpawnBox-Release-Corridor-Still-w3-v1	PushT-DROID-ZeroGrav-OnContact-v3
max_episode_steps	300	425	900	900	300	900	400	550
ent_coef	0.0	0.0	0.0	0.002	0.001	0.0	0.0	0.002
num_envs (train)	1024	1024	1024	2048	4096	1024	4096	4096
total_timesteps	50M	50M	50M	80M (checkpoint from an in-progress run, not final)	50M	50M	50M	100M
seed	1	1	1	1	1	1	1	2 (of 2, see the PushT section)

PullCube's ent_coef=0.001 (vs. 0.0 for the other three fixed-horizon tasks) reflects an empirical A/B on this task specifically: the ent_coef=0.001 run outperformed a matched ent_coef=0.0 run — this env's own containment box requires a released side wall (see "Env-specific note" below), and higher entropy exploration made a measurable difference here where it didn't for PushCube/LiftPeg/PickCube. num_envs=4096 (vs. 1024) reflects the batch-size convention used for this particular collection round, not a task-specific requirement.

Only PickCube's recipe was hand-tuned beyond the ManiSkill baseline (per the collaborator): max_episode_steps=900 (a arm_action_bound-bounded policy needs ~17s to complete the task) and gamma=0.95 instead of the baseline's 0.8 (0.8 gives an effective horizon of ~5 steps — far too short to carry reward across a 260+-step episode once actions are bounded). The other three tasks reuse that same gamma/horizon-scaling logic per-task.

PickSingleYCB's ent_coef=0.002 and total_timesteps=80M (vs. 0.0/50M for the other three) simply reflect that it's the only PickSingleYCB DROID run so far — no ent_coef A/B sweep was run for it like the other three, and its checkpoint is from a run still training toward 80M total steps at collection time (not a finished/final checkpoint).

Collection QA filtering (why episode counts differ from raw success rate)

Every kept episode passes ep_ever_succeeded (stock ManiSkill success flag true at some point) plus a per-task combination of additional checks, because the stock success condition alone can be satisfied by a transient, physically-unrepresentative moment (object slides through the target at speed, or briefly touches a grasp/placement condition without holding it):

Task	Filters applied	Why
PushCube	--require-success-at-end	Release-and-settle task; literal end-of-episode success check worked fine here (~66% of raw successes survived)
LiftPeg	--min-stable-hold-frames 15 (+ default settle-velocity check)	Release-and-settle task; --require-success-at-end proved too strict (~12% survival — many genuine successes drift by the literal last raw frame) — replaced with requiring the final success run last ≥15 frames wherever it falls
PickCube	--settle-vel-threshold 0 --require-grasp-at-success --min-stable-hold-frames 15	Grasp-and-hold task: the default settled-velocity check is the WRONG invariant for a held (still-moving-with-the-arm) object — disabled in favor of grasp-maintained + the same stable-hold-length check
PickSingleYCB	same as PickCube	Same grasp-and-hold success-condition family
PullCube	--require-success-at-end	Release-and-settle task (cube pulled into a goal region, arm releases); same rationale as PushCube — attempted 3,328, raw successes 2,168 (65.1%), dropped 198 for never settling + 642 for not still in the success state on the final step, kept 1,328 before the exact-1000 trim (survival 61.3%)
PlaceSphere	--full-horizon --min-stable-hold-frames 15 (+ default settle check), then placement QA (sphere on the bin for the whole kept tail)	Release-and-settle task; see the PlaceSphere section. Attempted 2,304, raw successes 2,215 (96.1%), 222 dropped by the collector (214 short final hold, 8 early ejection); 1,100 collected; placement QA kept 1,096/1,100; the first 1000 are shipped. First-success trim
RollBall	--full-horizon --settle-vel-threshold 0 --ejection-vel-threshold 1.0, then post-contact stillness QA (see the RollBall section)	Strike-and-roll task. Attempted 12,544, raw successes 2,827 (22.5%), 145 dropped by the ejection check; 1,400 collected; stillness QA kept 1,010/1,400; the first 1000 are shipped. First-success trim
PushT	--full-horizon --min-stable-hold-frames 0 (+ default settle check), then tail-coverage QA (see the PushT section)	Push task, success flickers. Attempted 4,352, raw successes 3,480 (80.0%), 58 dropped for early ejection + 189 never settling; 1,650 collected; tail-coverage QA kept 1,256/1,650; the first 1000 are shipped. First-success trim

--min-stable-hold-frames N requires the episode's final contiguous success run to last at least N frames wherever it occurs in the episode, rather than literally on the last raw-collected frame — verified empirically (via a direct probe) to still guarantee every kept episode survives the 15-frame hold-trim below ending on success=True, without requiring the much stricter (and here, much lower-yield) literal-final-frame condition.

PlaceSphere (added 2026-09-30): what differs from the other tasks

PlaceSphere was collected after the physics fixes found on 2026-09-28, with a different trimming rule, so its episodes are not produced exactly like the other tasks':

• First-success trim. Episodes end at the first success step + 15 frames, instead of the start of the final success run + 15 frames used for every other task. On PlaceSphere, success can flicker for a long time after placement. The stock success test requires the sphere to be static, and the sphere keeps spinning under the open gripper. The final-run trim would therefore ship long hover segments. The code is the opt-in --trim-mode first-success in scripts/h5_to_lerobot.py.
• Placement QA. On top of the collector's filters (--full-horizon --min-stable-hold-frames 15 + the default settle check), an episode is kept only if the stock is_obj_on_bin test holds for every frame of the 15-frame tail after first success. It was applied identically to both gravity arms, and the first 1000 passing episodes in collection order are shipped.
• Stock physics. The containment walls have their own material, and the sphere, bin and arm links use ManiSkill's stock restitution 0.0 / friction 0.3. The five earlier tasks have restitution 0.9 (see "Known issues"). Wall release uses a corrected mechanism: the wall moves only on the grasp-latch substep, and walls are restored on reset.
• Reward. This is stock ManiSkill PlaceSphere, with one DROID change: the ungrasp term reads the Robotiq gripper's opening instead of Panda finger joints, applied as a delta. ManiSkill's stock reward pays the robot for staying static after placing, and has no term for retracting. The experts therefore hover with the gripper open while the sphere settles.
• Settle time before success. Success requires the sphere to be static. In the ZG arm the sphere keeps spinning under the open fingers for ~12 s after release (median 178 steps), so ZG episodes are longer. The median first-success step is 355 (ZG) vs. 207 (regular gravity).

RollBall (added 2026-10-01): what differs from the other tasks

RollBall (strike a ball so it rolls into a goal region down a corridor) was collected after the physics fixes, with its own expert recipe and an extra QA filter on arm motion:

• Environment. The ball starts inside a small spawn box. In the ZG arm it floats with gravity cancelled until any robot link touches it (or the gripper comes within 5 cm). The spawn-box walls are then released and gravity switches on. The regular-gravity arm uses the same spawn box, released on the same contact trigger, with gravity on throughout. Outer corridor walls are never released. Stock materials (restitution 0.0) and the corrected release mechanism, as for PlaceSphere.
• Expert recipe. gamma=0.99 (not 0.95), num_steps=16, 8 update epochs, 32 minibatches, action_smoothness=1.0 with a 3000-step penalty warm-up, and a post-contact arm-motion cost (still_w=3): after contact, once the gripper is more than 8 cm from the ball, the expert pays for joint speed and action size. Without it, the experts kept moving the arm after striking the ball.
• Checkpoint: final_ckpt.pt, not best_ckpt.pt. For the ZG arm the final checkpoint beat the best one on both success (0.219 vs 0.180) and post-contact stillness (73% vs 67% of episodes), in a 256-env deterministic offline eval. The final checkpoint is used for both arms, to keep the pair symmetric.
• Post-contact stillness QA. An episode is kept only if the arm's mean joint speed (max over the 7 arm joints of |qvel|), from contact to the end of the shipped clip (first success + 15), is below 0.2 rad/s. Contact is the recorded spawn-box release step. It is computed from the recorded states only and applied identically to both arms. Pass rates are 72.1% (ZG) vs 93.8% (regular gravity), so the filter removes more of the ZG expert's behaviour. Within the shipped episodes the arms are similar (median 0.178 vs 0.172 rad/s).
• No settle check, first-success trim. The ball rolls through the goal rather than stopping, so the collector's settle-velocity check is off (--settle-vel-threshold 0). Episodes end at the first success + 15 frames (--trim-mode first-success).
• Timing. The median contact step is 70 (ZG) vs 74 (regular gravity). The median first-success step is 250 vs 168, and the mean peak ball height is 0.262 m vs 0.224 m.
• Open observation. The regular-gravity expert succeeded on 41.1% of collection attempts, above its 0.355 offline eval. The ZG expert matched its eval (22.5% vs 0.219). This has not been diagnosed.

PushT (added 2026-10-02): what differs from the other tasks

PushT (push a T-shaped block so it covers a T-shaped target outline: success = at least 90% overlap) was collected after the physics fixes, on a third version of its environments (-v3):

• Environment fixes (v2 and v3). The contact latch checks every arm and gripper link, not only the gripper base; the robot links get their own physics material (they used to inherit the walls' restitution 0.9); all of the T's collision shapes are blocked by the walls (before, its stem passed through them). v3 adds a guard so contact forces left over from the previous episode can't trigger the latch right after a reset; spawn clearance is checked against the arm's home pose; and the walls are 3 cm further out so a spawned T can never overlap one (the spawn region is unchanged).
• Experts. PPO, gamma=0.99, ent_coef=0.002, 4096 envs, 550-step horizon, 100M steps, two seeds per arm. The seed was chosen by a rule fixed before the results were seen, applied identically to both arms: use the seed whose best_ckpt.pt scores higher in a 256-env deterministic offline eval (ZG 0.777 vs 0.836 → seed 2; regular gravity 0.945 vs 0.863 → seed 1).
• First-success trim + tail-coverage QA. The experts push the T through the goal rather than stopping on it (median success streak: 2 frames), so the final-streak trim used by most tasks would keep almost nothing. Episodes end at the first success + 15 frames, and an episode is kept only if the T still covers at least 75% of the target for all 15 of those frames (success needs 90%). The rule is identical for both arms and keeps a similar share of each (76.1% ZG, 78.2% regular gravity).
• Timing. Median first-success step 163 (ZG) vs 190 (regular gravity); mean peak T height 0.207 m vs 0.134 m. ZG experts learned faster (their success took off at ~14M steps, vs ~29M or later for regular gravity, where one seed was stuck for ~50M steps translating the T without rotating it).

Hold-segment trimming

Every episode is collected to its full horizon (--full-horizon), then trimmed post-hoc to (start of the final contiguous success run) + 15 frames (~1s at 15Hz) — enough real frames behind the final action chunk for a 15-step chunked policy (e.g. pi0.5) to see genuine stop-and-hold behavior, without shipping the full multi-hundred-frame post-success tail. Trimming is computed from the FINAL success run specifically (not the first-ever success), since a --full-horizon episode can flicker (an early transient success that later drifts away isn't the one that holds) — this also composes correctly with --min-stable-hold-frames above.

Differences from zerograv-manipulation-trajectories-state-v0

• Recorded-action bug fixed: v0's actions were the policy's raw, unclipped output (RecordEpisode.step() logs the action before the env's own arm_action_bound clamp runs internally) even though the physically-simulated motion was correctly clamped. This round clamps the action before stepping, so recorded actions now match physics exactly (verified: max |arm action| == arm_action_bound across every episode).
• Hold-segment trimming added (see above) — v0 shipped the full post-success tail uncut.
• Corrected, task-appropriate QA filtering (see above) — v0 used --require-success-at-end uniformly for all tasks, which this round found to be the wrong/too-strict check for LiftPeg and the grasp-and-hold tasks.
• Raw HDF5 + provenance shipped: trajectory.h5/trajectory.json (full per-episode env_states, every simulation step — not just the trimmed frames that ship in the parquet) plus provenance.json and env_code/ (a snapshot of the exact env/training code at collection time) are included alongside the LeRobot data//meta/ directories, for independent reconstruction/cross-checking.
• One new task: PickSingleYCB (grasp-and-place on a randomized YCB object), not present in v0.

Key Notes

The collection-expert checkpoints are published as W&B model artifacts in the project collection-models, one run per task (linked in the stats table above). Each run's artifact contains the exact checkpoint used for collection (best_ckpt.pt; RollBall: final_ckpt.pt); its sha256 matches provenance.json. The artifact also has a config.json (env id, control mode, arm action bound, horizon, obs/action dims, training recipe) and a standalone agent.py (PPO actor-critic; load with Agent(n_obs, n_act) + load_state_dict, act with get_action(obs, deterministic=True)). Each stats-table row links two runs: training run is the original PPO training run, with the full training curves (eval success, return, losses), in the W&B projects Post_Recovery_Experiments (PushCube, LiftPeg, PickCube, PickSingleYCB) and domain-randomization-baseline (PullCube, PlaceSphere, RollBall, PushT). model is the collection-models run holding the checkpoint artifact. best_ckpt.pt is the best-eval checkpoint, so it usually comes from partway along the training curve, not its end.

observation.state layout is unchanged from v0/earlier exports for the same task (verified directly: PushCube 43-dim, LiftPeg 40-dim, PickCube 50-dim, all identical to the prior published exports) — this pipeline never transforms the state vector itself, only which frames are kept. PullCube's observation.state is 43-dim, matching PushCube's own single-object layout. RollBall's observation.state is 52-dim, PushT's 43-dim.

Env-specific note (PullCube): PullCube-DROID-ZeroGrav-OnContact-Confined-v1 uses a narrow containment box around the cube's pre-contact floating phase, with one side wall (the one between the box and the goal region) released the instant that env's gravity-cancellation latch fires — without this, the cube is physically unable to reach the goal at all, regardless of policy quality (same root cause and fix pattern as this project's PlaceSphere task; see environment_design/PlaceSphere.md). Mean peak object height 0.360m; median steps to first success 188 (of 300).

Known issues (found 2026-09-28)

These were found after publication while debugging later tasks. The data has not been re-collected. Fixes exist in the code for future collections. The exact code used for collection is still in each task's env_code/.

1. PullCube: slowed cube dynamics and inactive confinement. PullCube-DROID-ZeroGrav-OnContact-Confined-v1 releases one containment wall when the robot first contacts the cube. The release code had two bugs:

   • 1/5-speed dynamics. At every physics substep it pushed stale state back into the simulator, so free-moving objects advanced only one of every five substeps. Measured with the same code: free-fall speed after 5 control steps was 0.49 m/s instead of 2.45 m/s. This applied for the whole episode, not only after the release. The cube in these episodes therefore moves about 5x more sluggishly than real physics. Pulling behaviour is still present and successful, but the object dynamics are not physical.
   • Confinement not restored. The released wall was never put back when an environment reset, so after each parallel environment's first contact the pull-only confinement was absent for all its later episodes. The intended guarantee that a cube can only reach the goal by being pulled does not hold for most collected episodes.

   The regular-gravity PullCube twin in the companion dataset uses a different environment without wall release and is not affected. The PullCube ZG / regular-gravity pair therefore differs in more than gravity.

2. The five tasks collected before 2026-09-30 (not PlaceSphere): object restitution 0.90 (stock ManiSkill: 0.00). Manipulated objects share a physics material with the invisible containment walls, and the wall restitution (0.9) overwrites the object's value. In these zero-gravity episodes the effect is small: floating contacts are mostly slower than PhysX's 2 m/s bounce threshold, below which restitution has no effect, and bouncing off the walls is intended. See the companion regular-gravity card for where it does matter.

3. PickSingleYCB (found 2026-10-01): parts of an object could pass through the containment walls. The walls block only collision shapes in the object's collision group, and the collection code put only each object's first collision shape in that group. YCB objects are built from several convex hulls (1 to 21 per object). In a sample of 8 objects, 27 of 35 hulls were not blocked by the walls, so those parts could pass through a wall while the rest of the object was held back. Single-shape objects (every other task in this release: cube, peg, sphere, ball) are not affected. The bug is fixed in the code for future collections; how often it occurred in the shipped YCB episodes has not been measured yet.