LeeHakHo/mimicgen_aux_data

MimicGen ID90 / OOD — 12-task generalization suite

Generated demonstrations and fixed evaluation scenes for a study of where an auxiliary object-state head buys out-of-distribution generalization.

Each task is generated on a rotation-restricted ID90 training distribution and evaluated on single-factor OOD splits, so a drop can be attributed to position or to rotation rather than to both at once — three coarse rungs, plus a five-level graded ladder on each axis.

split	placement of the primary object
ID90 (training)	yaw restricted to a 90° window on the object's own declared center; x/y untouched
OOD_POS	x/y widened 1.67× about their own centers, yaw held at ID
OOD_YAW	yaw widened to a 180° window, x/y held at ID
OOD_BOTH	both

Each single-factor axis is also resolved as a five-level graded ladder, so a task's degradation reads as a curve against distance from training rather than as one number:

split	placement of the primary object
OOD_POS_L1..L5	x/y half-width × 1.2, 1.4, 1.6, 1.8, 2.0 about the ID center; yaw held at ID
OOD_YAW_L1..L5	x/y held at ID; yaw drawn from the k-th of five disjoint bands walking outward from the edge of the ID window to the far side of the object's orientation orbit

The two ladders differ in kind on purpose. The position levels are nested — that is what a scale factor means. The rotation levels are disjoint: a merely widened yaw window keeps re-drawing the ID orientations it contains, so L5 would still be mostly ID. With bands, every episode at level k is at least k−1 steps outside anything the policy trained on, and L5 sits at the far end of the orbit (180° for an asymmetric object, 45° for a C4 cube).

The bands are one-sided (they walk in +yaw only). |Δyaw| ∈ [a, b] on both sides is not expressible as the single (lo, hi) interval robosuite's UniformRandomSampler takes, and the only symmetric single interval outside a window is the whole outside — which is nested, not disjoint. So an L5 number means "180° away on the + side", not an average over both directions.

The axis a rung does not move equals ID exactly. Every object starts from the same ID spec, so POS and YAW are separately attributable. Only the primary (manipulated) object leaves the distribution — target objects keep the ID distribution, so the ID→OOD drop is attributable to the manipulated object alone.

Contents

id90/<task>/demo.hdf5                    200 successful demos, 84x84
eval_scenes/<task>/scenes_id.hdf5        60 held-out ID initial states
eval_scenes/<task>/scenes_ood_pos.hdf5   20 initial states
eval_scenes/<task>/scenes_ood_yaw.hdf5   20
eval_scenes/<task>/scenes_ood_both.hdf5  20
eval_scenes/<task>/scenes_pos_l1.hdf5    20   graded position ladder, ×1.2
  ...                    _l5                  ×2.0
eval_scenes/<task>/scenes_yaw_l1.hdf5    20   graded rotation ladder, band 1 of 5
  ...                    _l5                  band 5 (far side of the orbit)

demo.hdf5 carries agentview and eye_in_hand images at 84×84 plus obs/object, so it trains directly with no dataset_states_to_obs render pass.

The eval scenes are sim states only (216 MB total). They are shipped rather than regenerated because regeneration is only bit-identical under the same robosuite/MuJoCo/numpy stack — a numpy 1.x→2.x change alone has been enough to move evaluation results in this line of work. Using these files makes numbers from different machines comparable.

Tasks and datagen success rate

task	datagen SR	task	datagen SR
stack_d1	76%	mug_cleanup_d1	10%
square_d2	52%	coffee_d2	26%
threading_d0	50%	kitchen_d1	44%
stack_three_d1	41%	pick_place_d0	28%
three_piece_assembly_d0	34%	nut_assembly_d0	65%
hammer_cleanup_d1	42%	coffee_preparation_d1	21%

Every split holds 200 successful demos regardless of the rate. Note that the training set is therefore conditioned on datagen succeeding, while ID evaluation samples unconditionally — on the low-rate tasks the demos may be biased toward easier poses.

Two rungs are degenerate by construction and equal ID: coffee_d2's OOD_YAW (coffee_pod is a body of revolution) and pick_place's OOD_POS (its x/y range is the bin interior). This holds level by level on the graded ladder too — every Coffee_D2_OOD_YAW_L* and every PickPlace_D0_OOD_POS_L* equals ID. Reporting either as OOD is a real hazard: it is what put nine wrong cells in this study's first results table, where half an ID number was averaged in as an OOD one.

The yaw axis is not comparable across tasks — read the object's native orientation orbit before reading an OOD_YAW number. stack_d1 and stack_three_d1 use exact cubes (C4 about z, yaw_fold = 4), so their orbit is 90°, not 360°: their windows are divided by 4 to keep the ID rung at the same fraction of the orbit every other task gets (ID 22.5°, OOD 45°). Without that fold the 90° ID window already covered the whole orbit and the yaw rung measured nothing. threading_d0's needle and kitchen_d1's pot are 60° natively, so their ID→OOD_YAW is a 3× jump where every other task's is 2×.

hammer_cleanup_d1 uses HammerCleanup_FixedHead_Yaw45_Spawn25 as its ID env rather than the generic D1-based one: D1's hammer spawn is 4× wider in x and its drawer moves, which dropped datagen success to 4%. Two things are specific to it:

• The head direction is pinned. HammerObject.init_quat is a property that flips a coin on every access, so the head was drawn independently of the yaw window — and redrawn on every read, meaning the declared bounds and the sampled placement disagreed. Measured over 200 demos of the earlier Yaw45_Spawn12 version, the heading was bimodal (121 episodes in one lobe, 79 in the lobe 180° away): the declared ±45° band was really two 90° lobes. FixedHeadHammerObject pins it; the band stays at ±45°.
• The spawn is 2.5× the D0 half-width (0.20 × 0.175 m). At Spawn12 this was the easiest of the twelve tasks — its spawn area was 0.0080 m² against square_d2's 0.221 — and position is the axis policies actually fail on. The drawer stays pinned in both ID and OOD, since un-pinning it also moves the place target, which is what the single-factor ladder exists to separate.

nut_assembly_d0 needed a fix of a different kind: NutAssembly_D0 passed no placement_initializer, so robosuite built its own from bin geometry with rotation=None, and the bounds dict the ID90/OOD mixins rewrite was read by nothing — all four rungs sampled identically. It now builds a SequentialCompositeSampler from that dict the way Square_D0 already does.

Revisions

2026-09-02 — four tasks regenerated, graded ladder added. If you pulled before this date, id90/ and eval_scenes/ for those four tasks now hold a different distribution under the same paths. Results across the two versions are not comparable.

• nut_assembly_d0 — the ID90 restriction was never reaching the sampler. NutAssembly_D0 passed no placement_initializer, so robosuite built its own with rotation=None and the bounds dict the mixins rewrite was read by nothing: the training set carried full-circle nut yaw and all four rungs sampled identically. Datagen SR 45% → 65%.
• stack_d1, stack_three_d1 — the cubes are exact cubes (C4 about z), so the orientation orbit is 90°, not 360°, and the 90° ID window already covered all of it: OOD_YAW was resampling the same physical scenes. Windows are now divided by yaw_fold = 4 (ID 22.5°, OOD 45°). SR 75% → 76% and 46% → 41%.
• hammer_cleanup_d1 — ID env HammerCleanup_Yaw45_Spawn12 → HammerCleanup_FixedHead_Yaw45_Spawn25: the head direction is pinned (it was redrawn by a coin flip on every access) and the spawn widened to 2.5× the D0 half-width, since at Spawn12 this was by far the easiest task. SR 45% → 42%.
• New for all twelve tasks: eval_scenes/<task>/scenes_{pos,yaw}_l1..l5.hdf5, the five-level graded ladder (120 files).

The other eight tasks' demo.hdf5 and four rung scene files are byte-identical to the first release and were not re-uploaded.

2026-08-31 — first release. Twelve tasks, 200 demos each, four rungs of eval scenes.

Reproducing

The environments and the pipeline that generated this:

• https://github.com/LeeHakHo/mimicgen_aux — the <Base>_ID90 / _OOD_* env ladder plus the datagen → training → eval scripts
• https://github.com/LeeHakHo/Robomimic_Async branch feature/id90-12task — the policy-side code (aux heads, multi-object target blocks, fixed-scene evaluation)

Provenance and license

Generated with MimicGen from the NVlabs source demonstrations, in robosuite. It is a derivative of those datasets and inherits their non-commercial terms; released here under CC-BY-NC-SA-4.0 to match. Cite MimicGen and robosuite if you use it.