Skip to content

About

Isaac Sim / Isaac Lab benchmark for Alex humanoid door manipulation, scripted rollouts, dataset export, and action-space evaluation.

Topics

Resources

Stars

0 stars

Watchers

0 watching

Forks

Repository files navigation

AlexDoor-XAS

AlexDoor-XAS studies how action representation affects contact-rich humanoid manipulation: A1–A4 × ACT/Diffusion on held-out push doors, using fixed-base Purdue Alex003, WSG32/UMI v1 and head ZED RGB-D.

Current scope

  • Runtime and corpus: Purdue control/sensing and Phase 5 are complete. The frozen corpus contains 32 qualified doors: 19 train, six development, seven test.
  • Perception: maintained static GroundingDINO/SAM3 scans and an experimental CAD-free Point2Pose tracker. The selected SAM3 recipe is explicit; material identity, 150 ms freshness and loaded contact remain unqualified.
  • Learning: B1 observed-input datasets, ACT/Diffusion models, normalization and A1–A4 execution adapters are maintained. Qualified perception integration and physical rollout validation remain pending. No matched B1 policy dataset or learned-policy result exists; the sealed test remains closed.

The project status owns current results, limits and next work. No repository command controls physical hardware.

Installation

Use Python 3.11+. The supported workstation uses Isaac Sim 6.0.1, Isaac Lab release/3.0.0-beta2 and the external Alex package. Isaac, Alex, PyTorch, Warp and CUDA are supplied by that runtime; ordinary Python dependencies are declared in pyproject.toml.

/home/pacquadr/IsaacLab/isaaclab.sh -p -m pip install -e /home/pacquadr/Desktop/Alex
/home/pacquadr/IsaacLab/isaaclab.sh -p -m pip install -e '.[dev,diffusion]'
/home/pacquadr/IsaacLab/isaaclab.sh -p scripts/check_env.py

Robot/camera assets come from the external Alex package. Prepared door payloads, recordings and model weights remain local and ignored by Git. Perception workers use isolated dependencies; follow model setup.

Verification

ruff check src scripts tests
ruff format --check src scripts tests
/home/pacquadr/IsaacLab/isaaclab.sh -p -m pytest -q tests/unit
/home/pacquadr/IsaacLab/isaaclab.sh -p -m pytest -q tests/gpu
/home/pacquadr/IsaacLab/isaaclab.sh -p -m pytest -q tests/integration

GitHub Actions runs lint/format, unit tests on Python 3.11/3.12, wheel/sdist builds and an installed-wheel import check. Unit tests use CPU tensors without Isaac, model downloads or local payloads. CUDA model/storage regressions and integration with external Isaac/Alex/native sources run locally; GPU tests explicitly skip when CUDA is unavailable. A green public CI does not establish simulator, perception or physical qualification.

The complete Purdue runtime gate uses synthetic fixtures:

/home/pacquadr/IsaacLab/isaaclab.sh -p scripts/verify_purdue_runtime.py --viz none --device cuda:0
/home/pacquadr/IsaacLab/isaaclab.sh -p scripts/verify_door_corpus.py

Runtime reports default to ~/.cache/alexdoor-xas/verification/. Corpus verification checks tracked membership and references; --evidence-root also checks local expert reports. See Phase 5 for prepared-door intake and qualification.

Run bounded frozen perception inference with a fresh output directory:

/home/pacquadr/IsaacLab/isaaclab.sh -p scripts/perception.py smoke \
  --output outputs/generated/b1/perception/NEW_SMOKE

The model setup guide documents scan, Point2Pose smoke, operational replay, live and serial offline commands. Live diagnostics do not need recordings. Successful execution and scientific quality are separate outcomes; smokes do not admit motion or qualify a policy provider.

Organization

Path Responsibility
src/alexdoor_xas/ Runtime, action math, preparation, recording, B1 data and policies.
src/alexdoor_xas/perception/ Shared contracts/geometry; point2pose, sam3 and diagnostics subpackages.
scripts/, configs/ Supported entry points and common runtime/experimental settings.
tests/unit, tests/gpu, tests/integration Checks separated by execution requirements.
datasets/doors/b1/ Frozen corpus, canonical records and ignored prepared payloads.
knowledge/ User-owned raw research and canonical technical wiki.
datasets/recordings/ Local reusable recordings and policy data.
outputs/generated/, outputs/evidence/ New runs and preserved results, respectively.
models/ Entirely ignored local weights and isolated runtimes; setup lives in the wiki.

B0 workflows, numerical phase2.v2 data APIs and standalone v3 checkpoint loaders are retired; historical source remains at 097d578. B1 formats and observation/ action ordering are unchanged. New perception imports are documented in Shared Door Perception.

Start at the wiki index, architecture, data contracts and policy stack.

License

This repository is proprietary. No license grant is provided unless a separate license file or written agreement states otherwise. Door assets retain their recorded third-party terms: 29 redistributable, two local-only and one private/noncommercial.

About

Isaac Sim / Isaac Lab benchmark for Alex humanoid door manipulation, scripted rollouts, dataset export, and action-space evaluation.

Topics

Resources

Stars

0 stars

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages