SII-EricSanchez/act2sim-data

Act2Sim v0.6.0 public task data

This revision recalibrates the rigid tasks under the rigid physical-scoring rules: part-owner correspondence, observation-noise eligibility and a reported single-observation control (stored as ceiling).

• Drawer and peg are evaluable at Levels 1–3.
• Peg uses a new physics-sensitive probe panel that tests the peg's mass.
• Cloth ships its interaction assets only. This revision publishes no calibrated cloth evaluation.

Scores from v0.5.0 and v0.6.0 are not comparable. Compare results only within one task and protocol identity.

v0.6.0 is an immutable dataset artifact revision; the research version remains v1.3. Core F/R amendments accumulate in the append-only algorithm record under the research-version policy. Routine engineering and documentation fixes do not add entries or create a new research release.

Act2Sim evaluates worlds built through code:

1. An agent interacts with a target world.
2. It writes a source submission describing its reconstructed world.
3. It evaluates that submission on the user's own machine.

The Python package supplies the API, simulation and evaluator. This dataset supplies the target assets, public robot assets, frozen evaluation data and perceptual metric weights. No hosted evaluator, agent upload, model-service account or benchmark-server compute is required.

Public target and evaluation files are downloadable. The level-specific API controls the information supplied to the agent; it is not a security sandbox against a user reading downloaded evaluator files. Researchers are responsible for keeping their reconstruction agent within the declared information boundary, and for disclosing any use of additional target information.

Tasks and levels

Task	Task revision	World	Robot configuration	Evaluation in this revision	Scoring device
drawer/open_stop	drawer-v1	Passive RoboCasa drawer opening and stopping	One FR3 workcell arm	Levels 1–3	CPU
assembly/peg_insert	peg-v1	Factory peg and hole; the probes grasp, loosen and shake the peg	One FR3 workcell arm	Levels 1–3	Linux CUDA, logical cuda:0
cloth/unfold_shirt	shirt-v2	Coarse DexGarmentLab shirt manipulation	Two FR3 workcell arms	Interaction only	—

Level	Information provided through the agent API	Reported scores
1	RGB, depth and robot state, plus initial exterior geometry and noisy whole-object poses	Physical F
2	RGB, depth and robot state; initial target geometry and object poses are omitted	Physical F
3	The Level 2 information boundary, with appearance reconstruction additionally evaluated	Physical F and appearance R

The score columns apply to calibrated tasks only. The levels use the same underlying task physics and interaction implementation. Select a level with level=1, level=2 or level=3 in act2sim.open_task(...) and act2sim.evaluate(...), or with the --level argument.

Cloth can be opened, explored and reconstructed as before. Evaluating it with this revision fails before the submission is sealed, with an error naming the task and the data revision. The report status is runtime_error and carries no score.

Run locally

The code repository is currently private. The code and installation-reference links in this card require repository access; public dataset downloads are separate and do not require a Hugging Face token.

The complete scoring profile is Linux x86-64, Python 3.12 and the locked CUDA 12 dependency builds. Drawer uses CPU; peg uses CUDA.

A compatible NVIDIA driver, LLVM/OptiX native rendering libraries and sufficient local resources remain host prerequisites. On Linux, complete the native-library recipe before running doctor or the example below.

git clone https://github.com/EricSanchezok/act2sim.git
cd act2sim
git checkout f98b3e674109b8467a14def572264b82d93fea06
uv sync --locked --python 3.12 --group dev --group assets
uv run act2sim doctor
uv run python examples/local_agent.py --task drawer/open_stop --level 1 --device cpu --output results/drawer-l1

Package 0.6.0 resolves the immutable v0.6.0 dataset tag by default. The recommended consumer revision is f98b3e674109b8467a14def572264b82d93fea06; it includes the merged calibration-control validation and reporting fixes. These fixes preserve the calibration digest below.

The executable example demonstrates interaction, reset, a source submission and local evaluation. Its fixed generic reconstruction is a workflow example, and its scores do not establish reconstruction quality. Replace that reconstruction with your own agent's source submission.

Use assembly/peg_insert with device="cuda:0". To select a different physical GPU, set CUDA_VISIBLE_DEVICES before starting Python and keep logical cuda:0.

Reproducing a v0.5.0 score needs the v0.5.0 code commit named in that revision's card. Package 0.6.0 rejects v0.5.0 calibrations with a calibration-identity error.

Automatic downloads and offline use

The package downloads the requested public files automatically. It resolves the selected Hub revision to an immutable commit and verifies manifest sizes and SHA-256 values. A token is not required. Downloads happen in stages:

1. Interaction fetches the target and robot assets.
2. Levels 1–2 evaluation adds the task's calibration files.
3. Level 3 adds the shared perceptual-model payload, about 2.2 GB.

HF_HOME selects the cache, and offline=True uses verified cached data only. ACT2SIM_DATA_ROOT selects an already materialized tree, and ACT2SIM_DATA_REPO / ACT2SIM_DATA_REVISION select a mirror or immutable revision.

Dataset layout and evidence

manifest.json
README.md
tasks/drawer/target/
tasks/peg/target/
tasks/shirt/target/
robots/fr3/
evaluation/drawer/
evaluation/peg/
calibration-inputs/
calibration-inputs-lineage.json
quality/cloth/
metrics/
native-runtime/linux-x86_64/

manifest.json declares:

• drawer and peg with evaluation_root and metrics_root;
• cloth without them, which makes it interaction-only.

Task	Frozen protocol	Sealed calibration	Ceiling control
Drawer	Protocol	Calibration	Ceiling
Peg	Protocol	Calibration	Ceiling

calibration-inputs/ keeps the v0.5.0 reference and prior artifacts, which this revision reuses. Its lineage file still binds them. calibration-inputs/peg-panel-v0.6.0/ adds the candidate generator outputs and the prior-only selection records of the new peg panel.

quality/cloth/ is the v0.5.0 cloth quality record. It describes the cloth target that is still published for interaction.

Frozen evaluation configuration

Both evaluated tasks use:

• the 16 published prior members of v0.5.0;
• the v0.5.0 target reference;
• eight held-out appearance views, plus the original public views (drawer three, peg three);
• correspondence part_owner_v1.

Task	Probes (eligible)	ID / OOD / critical	Events (eligible)	Event term
drawer/open_stop	30 (28)	10 / 10 / 10	90 (10)	events
assembly/peg_insert	29 (29)	9 / 10 / 10	29 (29)	events

Correspondence. Points vote for a submitted motion owner using nearby surface geometry within a size-relative radius. A strict majority is measured over the points that cast a vote, not all points in the target part. If an owner wins, every point of that rigid target part follows it. There is no minimum voting-coverage threshold, so sparse coverage can determine the whole part's motion; reports expose vote coverage and splits. Drawer parts use about 32 mm and peg parts about 3.4 mm. This rule does not establish resistance to all geometric exploits.

Eligibility. A probe enters F only if its prior spread is at least the spread produced by one Level 1 pose-noise observation (ρ ≥ 1). An event enters F only if two conditions hold:

• its prior frequency lies in [1/16, 15/16];
• its prior variance is at least the pose-noise variance.

The decision uses prior replays and pose-noise draws only. Excluded probes and events stay in reports as sanity checks.

Single-observation control (ceiling). Eight Level 1 pose-noise draws of the target reference are replayed and scored as ordinary submissions. This measured reference isolates the effect of one noisy pose observation with exact physics. It is reported without entering eligibility selection or score normalization. It is not an upper bound on agent F or a guarantee of agent performance; repeated observations can reduce pose uncertainty.

Task	Single-observation control F mean (min–max)	Template-baseline F, Level 1	Template-baseline R, Level 3
Drawer	0.775 (0.586–0.883)	−4.230	−0.606
Peg	0.808 (0.633–0.945)	−2.067	0.169

The new peg panel grasps the peg with the public qualification tape and lifts it. It then opens the gripper to a fixed aperture and either holds or shakes. Whether the peg slips depends on its mass. Each candidate batch was fixed before its own replay; a later fill batch followed the initial pilot results. The selection thresholds were fixed, and selection used a separate pilot prior cohort with pose-noise draws and a cross-GPU portability check on an RTX 4090 and an RTX 5090. Target trajectories were not used for panel selection.

F uses equal weighting of continuous and event skill when eligible events exist, with losses aggregated before prior normalization. With no eligible event, F uses continuous skill alone. R is one minus mean LPIPS divided by the calibrated gray-control mean LPIPS. Scores may be negative. Read the report's validity status alongside its numbers.

Source and numerical identity

The original publication identifies Act2Sim commit ff95a30381871b9d56755950ca048099a6268af8 as its numerical implementation. This remains historical provenance; current installation guidance selects the compatible consumer revision above. Calibration code SHA-256: 68a83cf8bf6a8cafb290c30d8c31c036f31d7a68f9fe4628d0d4954e064f0038.

The dataset tag remains v0.6.0 at content commit e42eceb84f5770039c594f5599eb57d2d279bd68. Manifest SHA-256: c868129779ed9c560522ca55f27a962fbaba8ee53771c491c22c735e62f76dae. The current card corrects installation guidance and scoring explanations after tagging; the historical card at the tag is unchanged, and all non-card files remain identical.

Maintainers must follow the code/data/documentation alignment procedure when either repository changes. Keep current consumer guidance separate from experiment provenance, preserve immutable data tags, and verify both PR outcomes.

Calibration host:

• Ubuntu 22.04 x86-64 with RTX 4090 GPUs and driver 570.153.02.
• Drawer was sealed on CPU with llvm_ad_rgb.
• Peg was sealed on CUDA with cuda_ad_rgb, using OptiX from the matching NVIDIA driver package.

Numerical compatibility uses:

• the calibration digest;
• the backend family;
• the required renderer, for Level 3;
• the pinned numerical dependency identities.

Licenses and provenance

Unchanged from v0.5.0. This collection retains multiple upstream licenses, and Act2Sim's MIT code license does not relicense the task assets, robot assets or pretrained models.

Asset	Upstream source and revision	License identified by the asset notice	Dataset notice
Drawer	RoboCasa, 3443cdf3dd20f9ae8ed41adba442059b506891d0; OpenDrawer episode 20250819/episode_000000	CC-BY-4.0	Drawer NOTICE
Peg/hole	IsaacLab Factory, fd4b00bbf07a667bdadac8a4d5bd8c64b5a57a4a	BSD-3-Clause	Peg NOTICE
Shirt	DexGarmentLab, 2ba4092676006bc98c257e7b822de39526fd9692	Apache-2.0, as identified in the upstream dataset card	Shirt NOTICE
Act2Sim FR3 robot pack	Newton assets, Franka/Panda source, f8fb7abcbeba2318814a74f3eeb02780ad7925d6	Apache-2.0	Robot NOTICE

metrics/NOTICE.json and metrics/weights.json retain the perceptual-model sources, revisions, hashes and license notes.

The optional native LLVM runtime attachment and its copyright notices are unchanged. NVIDIA driver and OptiX libraries are not redistributed.

Scientific limitations and intended use

• Peg tests mass, not friction.
  • Finger–peg friction is the larger of the finger's 1.0 and the peg's coefficient, so peg friction affects grasp slip only above 1.0.
  • Pushes of the standing peg were measured as noise-dominated and are not in the panel.
  • A good peg F therefore says little about inferred friction.
• Peg outcomes contain chaotic landings. A peg that slips lands upright or lying. Selection kept only candidates whose cross-GPU difference was at most half the pose-noise spread. Untested GPU models may still differ more.
• The single-observation control is not an agent ceiling. Its observed scores are below 1 for these draws; that measurement is neither a universal strict bound nor a limit on an agent that obtains repeated observations.
• Statistical coverage is limited. There are 16 prior members and four bases with eight draws each. Ranking stability is not established.
• Cloth is not evaluable in this revision. Its v0.5.0 quality findings still describe the published target.

Report the following when sharing results:

• task, level, data and protocol identities;
• runtime identity;
• reconstruction method;
• validity status.