The Dataset Viewer has been disabled on this dataset.

Act2Sim Public Task Data (v0.7.0)

Versioned benchmark assets, sealed evaluation protocols, and perceptual metric weights for Act2Sim (0.7.0).

Note: This repository is a manifest-indexed asset store consumed automatically by the act2sim Python package—not a datasets.load_dataset(...) image-classification dataset.


Code vs. Data Boundary

Repository Responsibility
EricSanchezok/act2sim Python SDK (open_task, Session, evaluate), Newton/MuJoCo simulation runtime, CLI, and benchmark documentation
SII-EricSanchez/act2sim-data (this repo) manifest.json (SHA-256 verified), target scenes (tasks/), FR3 robot pack (robots/), sealed calibration protocols (evaluation/), perceptual weights (metrics/), and provenance records

What Changed in v0.7.0

  • drawer/open_stop (drawer-v1): 72-probe panel, trajectory-only physical scoring (continuous_only, no event term).
  • assembly/peg_insert (peg-v2): Robot fingers now read the peg's own friction with a stiffened mimic coupling (equality_timeconst_substeps=2); 60-probe loosening panel; trajectory-only physical scoring (continuous_only) sensitive to both mass and friction.
  • cloth/unfold_shirt (shirt-v2): Unchanged from v0.6.0 (interaction-only; no calibrated evaluation in v0.7.0).

Drawer and peg scores from v0.6.0 and v0.7.0 are not comparable. Compare results only within matching task and protocol identities.


Tasks & Information Levels

Task Revision World Robots Evaluation (v0.7.0) Scoring Device
drawer/open_stop drawer-v1 Passive articulated drawer 1× FR3 Levels 1–3 CPU
assembly/peg_insert peg-v2 8 mm peg-and-hole assembly 1× FR3 Levels 1–3 Linux CUDA (cuda:0)
cloth/unfold_shirt shirt-v2 Deformable shirt 2× FR3 Interaction only —
Level Observations Provided to Agent Reported Metrics
1 RGB-D, robot proprioception, exterior geometry, noisy whole-object poses Physical F
2 RGB-D and robot proprioception (no object geometry or poses) Physical F
3 Same as Level 2, with appearance reconstruction evaluated Physical F + Appearance R

Quickstart (Automatic Download)

You do not need to clone this repository manually or pass a Hugging Face token. act2sim downloads and SHA-256-verifies only the files needed for each stage:

  1. Interaction: target & FR3 robot assets
  2. Levels 1–2 Evaluation: task calibration protocol & reference trajectories
  3. Level 3 Evaluation: shared perceptual-model weights (~2.2 GB)
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

Environment overrides: HF_HOME (cache path), ACT2SIM_DATA_ROOT (local unpacked snapshot), ACT2SIM_DATA_REPO / ACT2SIM_DATA_REVISION, and offline=True (verified cache only). On Linux, complete the native-library recipe first; for assembly/peg_insert, keep logical cuda:0 and select physical GPUs via CUDA_VISIBLE_DEVICES.


Repository Layout

manifest.json                     # Authoritative file inventory, sizes, and SHA-256 digests
tasks/{drawer,peg,shirt}/target/  # Public target scene assets and upstream NOTICE files
robots/fr3/                       # Shared FR3 robot pack
evaluation/{drawer,peg}/          # Frozen protocols, priors, ceiling controls, and views
metrics/                          # Shared perceptual metric weights (LPIPS / CLIP / DINO / LoRA)
calibration-inputs/               # Panel selection & calibration provenance records
calibration-inputs-lineage.json   # Lineage binding v0.5.0 priors/references to v0.7.0
quality/cloth/                    # v0.5.0 cloth target quality diagnostics
native-runtime/linux-x86_64/      # Optional LLVM runtime attachment and notices

Calibration & Validity Summary

Both evaluated rigid tasks use part_owner_v1 correspondence, 16 prior members from v0.5.0, 8 held-out views + 3 public views, and continuous_only F scoring (F = 1 - S / S_prior; negative scores preserved). R is 1 - mean_LPIPS / gray_control_mean_LPIPS.

Task Probes (Eligible) ID / OOD / Critical Ceiling F Mean (Min–Max) Template Baseline F (L1) Template Baseline R (L3)
drawer/open_stop 72 (70) 24 / 24 / 24 0.530 (0.029–0.779) −4.804 −0.606
assembly/peg_insert 60 (60) 20 / 20 / 20 0.673 (0.397–0.837) −7.032 0.169

Preregistered validity studies (spec 0013 for drawer, spec 0014 for peg):

  • Drawer (v0.7.0): Pairwise ordering 0.907 (95% CI 0.841–0.940), lowest per-class 0.879 (moving mass), half-panel 5th percentile 0.948.
  • Peg (v0.7.0, 60 probes): Pairwise ordering 0.972 (95% CI 0.965–0.981), lowest per-class 1.00, representation variants 1.00, half-panel 5th percentile 0.958, memorization check 1.00.

Source & Numerical Identity:

  • Code commit: e3cdf88 · Calibration code SHA-256: 38e6463b5a61903b071a650f5457a523ca6953ddfe01f264a2dc143cbef89142
  • Calibration host: Ubuntu 22.04 x86-64, RTX 4090 (driver 570.153.02), llvm_ad_rgb (drawer CPU) / cuda_ad_rgb + OptiX (peg CUDA).

Licenses & Provenance

This collection retains multiple upstream licenses; Act2Sim's MIT code license does not relicense task assets, robot assets, or pretrained models.

Asset Upstream Source & Revision License Dataset Notice
Drawer RoboCasa 3443cdf3dd20f9ae8ed41adba442059b506891d0 (OpenDrawer 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 Shirt NOTICE
FR3 robot pack Newton assets f8fb7abcbeba2318814a74f3eeb02780ad7925d6 Apache-2.0 Robot NOTICE

See metrics/NOTICE.json, metrics/weights.json, native-runtime/linux-x86_64/receipt.json, and native-runtime/linux-x86_64/COPYRIGHT for metric and runtime notices.


Scientific Limitations

  • Peg measures mass and friction through slip; contact geometry is weakly probed, high friction values above the closed-grip threshold cost little F, and chaotic landing outcomes (upright vs. lying) are filtered for RTX 4090 / 5090 consistency only.
  • Ceiling is below 1 by construction (single noisy Level 1 observation with exact physics; drawer ceiling spans 0.03–0.78).
  • Drawer events affect F only through trajectory distance; cloth is interaction-only in v0.7.0.
  • Statistical coverage is limited to 16 prior members, 4 bases × 8 draws, and synthetic confirmation candidates (6 drawer, 2 × 4 peg).
Downloads last month
1,351