File size: 18,545 Bytes
f5a5947 e20af7b f5a5947 e20af7b f5a5947 019a588 f5a5947 e20af7b 019a588 e20af7b c9cf034 019a588 e20af7b c9cf034 019a588 c9cf034 e20af7b c9cf034 019a588 e20af7b c9cf034 019a588 c9cf034 019a588 e20af7b 019a588 e20af7b 019a588 e20af7b 019a588 e20af7b 019a588 c9cf034 019a588 e20af7b 019a588 c9cf034 019a588 c9cf034 019a588 c9cf034 019a588 c9cf034 019a588 c9cf034 019a588 c9cf034 f5a5947 e20af7b f5a5947 c9cf034 f5a5947 019a588 c9cf034 019a588 c9cf034 019a588 e20af7b 019a588 e20af7b 019a588 e20af7b 019a588 e20af7b 019a588 e20af7b 019a588 e20af7b c9cf034 e20af7b 019a588 e20af7b 019a588 e20af7b 019a588 e20af7b 019a588 e20af7b 019a588 e20af7b 019a588 e20af7b 019a588 c9cf034 019a588 c9cf034 f5a5947 c9cf034 019a588 c9cf034 019a588 c9cf034 019a588 e20af7b c9cf034 e20af7b c9cf034 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 | ---
pretty_name: Act2Sim v0.7.0 public task data
license: other
license_name: mixed-upstream-licenses
license_link: https://huggingface.co/datasets/SII-EricSanchez/act2sim-data/blob/v0.7.0/README.md#licenses-and-provenance
task_categories:
- robotics
tags:
- act2sim
- physical-simulation
- scene-reconstruction
- robotics
---
# Act2Sim v0.7.0 public task data
This revision changes drawer scoring and the peg task. Cloth is unchanged from `v0.6.0`.
- **Drawer** uses a 72-probe panel and scores trajectories only: F has no event term. Both changes passed a sealed validity check against an independent behavior measure, where `v0.6.0` drawer scoring failed it.
- **Peg** is task revision `peg-v2`:
- The robot's fingers now read the peg's own friction and grip it with a stiff coupling.
- A new 60-probe panel loosens the grip so that the peg slips.
- F scores trajectories only and now responds to both the peg's mass and its friction.
- It passed the same sealed validity check twice; `peg-v1` had failed it.
- **Cloth** ships its interaction assets only. This revision publishes no calibrated cloth evaluation.
Drawer and peg scores from `v0.6.0` and `v0.7.0` are not comparable. Compare results only within one task and protocol identity.
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-v2` | 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 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](https://github.com/EricSanchezok/act2sim/blob/e3cdf88/docs/deployment.md#native-rendering-libraries) before running `doctor` or the example below.
```sh
git clone https://github.com/EricSanchezok/act2sim.git
cd act2sim
git checkout e3cdf88
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 `0.7.0` resolves this revision by default.
The [executable example](https://github.com/EricSanchezok/act2sim/blob/e3cdf88/examples/local_agent.py) 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 `v0.6.0` or `v0.5.0` score needs the code commit named in that revision's card. Package `0.7.0` rejects their 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
```text
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](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](evaluation/drawer/protocol.json) | [Calibration](evaluation/drawer/calibration.json) | [Ceiling](evaluation/drawer/ceiling.json) |
| Peg | [Protocol](evaluation/peg/protocol.json) | [Calibration](evaluation/peg/calibration.json) | [Ceiling](evaluation/peg/ceiling.json) |
`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-v0.7.0/` keeps the peg panel's records:
- the 270 candidates with their pilot cohort;
- the prior-only selection records and the selected panel;
- the record of the revision run that sealed it, and its comparison with the validity study's calibration.
- `calibration-inputs/drawer-panel-v0.7.0/` keeps the same records for the drawer 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` | 72 (70) | 24 / 24 / 24 | 216 (0) | `continuous_only` |
| `assembly/peg_insert` | 60 (60) | 20 / 20 / 20 | 60 (0) | `continuous_only` |
**Correspondence.** Every point of a rigid target part follows the submitted body that holds a strict majority of the part's points within a size-relative radius. Drawer parts use about 32 mm and peg parts about 3.4 mm. A submission that splits a part across bodies is judged by the majority owner, and the split is reported.
**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). The decision uses prior replays and pose-noise draws only. Excluded probes stay in reports as sanity checks.
**Both tasks score trajectories only.** Every event thresholds a quantity of its own probe's trajectory, such as how far the drawer opened or whether the peg is still held. The continuous term already scores that trajectory. In the validity study, a single flipped event moved a drawer score by 0.18–0.34. Both calibrations therefore declare their task continuous-only:
- every event is recorded as ineligible with the reason `continuous_only`;
- F = 1 − S/S_prior on the continuous term;
- events and their outcomes remain in reports as diagnostics.
**Drawer panel.** The 30 `v0.6.0` probes were kept, and 14 new probes per category were added by a prior-only rule. Candidates and the rule were frozen before any replay. Selection used a separate pilot prior cohort. The target was not used.
**Peg physics (`peg-v2`).**
- The FR3 fingers' sliding friction is 0.01 in the precision profile. MuJoCo combines two shapes' friction by the larger value, so finger–peg contact now uses the peg's own coefficient. In `peg-v1`, the fingers' 1.0 hid it.
- Only one finger is driven; the other follows through an equality constraint. That constraint is now stiffened to two solver substeps, so the grip does not yield.
- The probes close the fingers fully from the grasp waypoint on.
- The geometry, priors and robot effort limits are unchanged.
**Peg panel.** 270 loosening candidates were generated and replayed on a separate pilot prior cohort, never on the target. 174 passed the prior-only selection rule. The panel keeps the 20 candidates per category with the largest prior-to-noise spread ratio among them. Each probe grasps the peg with the public qualification tape and lifts it. It then opens the fingers to a fixed aperture and holds or shakes. Whether and how fast the peg slips depends on its mass and its friction.
**Ceiling control.** Eight pose-noise draws of the target reference, scored as submissions, give the F an agent with exact physics and one noisy Level 1 observation can reach. It is reported and is never used to normalize.
| Task | Ceiling F mean (min–max) | Template-baseline F, Level 1 | Template-baseline R, Level 3 |
| --- | --- | --- | --- |
| Drawer | 0.530 (0.029–0.779) | −4.804 | −0.606 |
| Peg | 0.673 (0.397–0.837) | −7.032 | 0.169 |
The drawer ceiling is lower than the `v0.6.0` value of 0.775 because the event term no longer lifts it. On the same 72 probes with the `v0.6.0` event rule, the ceiling was 0.747: the noisy observations matched the events almost perfectly (event skill 0.964), while their continuous skill was 0.530, as here. By category, the drawer ceiling means are 0.600 (id), 0.386 (ood) and 0.603 (critical). One ood draw scores below zero: the eligibility rule admits probes whose single-observation spread is as large as the prior spread, so one unlucky observation can score as badly as a typical prior member on such a probe.
By category, the peg ceiling means are 0.676 (id), 0.513 (ood) and 0.766 (critical).
The `v0.6.0` card reported a peg ceiling of 0.808 for a different panel. That value was measured with a multiworld replay defect, corrected in this code revision: every batched pose-noise draw started from the first draw's peg pose.
F aggregates losses before prior normalization. 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.
## Validity evidence
Drawer and peg scoring in this revision were checked by preregistered studies: [spec 0013](https://github.com/EricSanchezok/act2sim/blob/e3cdf88/docs/specs/0013-rigid-f-validity-round-2.md) for drawer and [spec 0014](https://github.com/EricSanchezok/act2sim/blob/e3cdf88/docs/specs/0014-peg-v2-validity.md) for peg.
- Instances share the drawer geometry and vary its hidden physical parameters.
- Candidate reconstructions carry known parameter errors.
- An independent behavior measure, built from actions outside the panel, ranks the candidates.
- The study then asks whether F ranks them the same way, stably and without rewarding memorized trajectories.
On six newly sealed instances, scored once after the rules were committed:
| Measure | `v0.7.0` drawer | `v0.6.0` drawer, same instances | Required |
| --- | --- | --- | --- |
| Pairs ordered as the behavior measure orders them | 0.907 (95% CI 0.841–0.940) | 0.894 | ≥ 0.90 |
| Lowest per-class ordering | 0.879 (moving mass) | — | ≥ 0.80 |
| Ranking agreement between half-panels, 5th percentile | 0.948 | 0.869 (minimum 0.779) | ≥ 0.90 |
Prior resampling and the memorization check also passed.
Peg was checked the same way twice, on four newly sealed instances each time:
1. The first confirmation used a 102-probe panel and passed every testable criterion (ordering 0.962, half-panel 5th percentile 0.931).
2. The released 60-probe panel was then chosen by a fixed rule and confirmed on four further instances. Its outcome was fixed in advance: had it failed, the 102-probe panel would have been released instead.
| Measure | `v0.7.0` peg, 60 probes | `peg-v1`, round 1 sealed instances | Required |
| --- | --- | --- | --- |
| Pairs ordered as the behavior measure orders them | 0.972 (95% CI 0.965–0.981) | 0.767 | ≥ 0.90 |
| Lowest per-class ordering (friction, mass, structure, memorizer) | 1.00 | — | ≥ 0.80 |
| Representation variants ordered correctly | 1.00 | — | ≥ 0.90 |
| Ranking agreement between half-panels, 5th percentile | 0.958 | not stable | ≥ 0.90 |
| Memorization check | 1.00 | — | ≥ 0.90 |
The release calibration reproduces the peg calibration of that confirmation exactly: reference, prior and ceiling trajectories and every probe's eligibility are identical.
## Source and numerical identity
The numerical implementation for this revision is [Act2Sim commit e3cdf88](https://github.com/EricSanchezok/act2sim/commit/e3cdf88). Calibration code SHA-256: `38e6463b5a61903b071a650f5457a523ca6953ddfe01f264a2dc143cbef89142`. Peg prior, pose-noise and ceiling replays ran as multiworld batches over four GPUs (`scripts/validation/batched.py`), which reproduces the single-world replay bit for bit.
**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](https://github.com/robocasa/robocasa), `3443cdf3dd20f9ae8ed41adba442059b506891d0`; OpenDrawer episode `20250819/episode_000000` | CC-BY-4.0 | [Drawer NOTICE](tasks/drawer/target/NOTICE.md) |
| Peg/hole | [IsaacLab Factory](https://github.com/isaac-sim/IsaacLab), `fd4b00bbf07a667bdadac8a4d5bd8c64b5a57a4a` | BSD-3-Clause | [Peg NOTICE](tasks/peg/target/NOTICE.md) |
| Shirt | [DexGarmentLab](https://huggingface.co/datasets/wayrise/DexGarmentLab), `2ba4092676006bc98c257e7b822de39526fd9692` | Apache-2.0, as identified in the upstream dataset card | [Shirt NOTICE](tasks/shirt/target/NOTICE.md) |
| Act2Sim FR3 robot pack | [Newton assets, Franka/Panda source](https://github.com/newton-physics/newton-assets/tree/f8fb7abcbeba2318814a74f3eeb02780ad7925d6/franka_emika_panda), `f8fb7abcbeba2318814a74f3eeb02780ad7925d6` | Apache-2.0 | [Robot NOTICE](robots/fr3/NOTICE.md) |
[metrics/NOTICE.json](metrics/NOTICE.json) and [metrics/weights.json](metrics/weights.json) retain the perceptual-model sources, revisions, hashes and license notes.
The optional [native LLVM runtime attachment](native-runtime/linux-x86_64/receipt.json) and its [copyright notices](native-runtime/linux-x86_64/COPYRIGHT) are unchanged. NVIDIA driver and OptiX libraries are not redistributed.
## Scientific limitations and intended use
- **Peg measures mass and friction through slip.**
- Every probe loosens a grasp, so peg F sees mass and friction as they affect slip.
- It sees contact geometry poorly. Tilt probes were dropped because they penalized equivalent representations of the peg.
- When the peg's friction is high, the fully closed grip already holds it, so overestimating a high friction costs little F.
- Pushes of the standing peg were measured as noise-dominated and are not in the panel.
- **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 ceiling is below 1 by construction.** It measures one noisy observation with exact physics, and it shows how much of the scale is reachable. The drawer ceiling varies widely between single observations (0.03–0.78).
- **Drawer critical events count only through trajectories.** A reconstruction that misses a critical event, such as the drawer reaching its stop, loses F only through the trajectory distance this causes.
- **Statistical coverage is limited.** There are 16 prior members and four bases with eight draws each. The validity studies used six drawer and twice four peg confirmation instances, all of one geometry per task. Their candidates were synthetic, with known errors, not real agent submissions.
- **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.
|