1,000 articulated door models for robotics simulation.
Browse · Review doors · Download the dataset · Release guide
DoorBench combines 30 door families with operators, locks, closers and configurable materials. It includes detailed model specifications, simulation exports, benchmark scenarios, Blender appearances, and reference motion for studying door interactions.
The robotics collection contains 985 doors. The 15 pet doors are a separate, downloadable supplement, excluded from evaluations and benchmark scores.
- Simulation: full, simple and minimal MJCF/URDF; full and canonical USD; shared OBJ/USDC hardware meshes.
- Appearance: 1,014 saved Blender renders, interchangeable finishes and lighting, CC0 textures, and available packed scenes.
- Door trajectories: native scripted-hand recordings with explicit outcomes and tracking errors. Experimental figure overlays are retained for diagnosis; natural human references are in development.
- Reproducibility: versioned public downloads, per-file checksums, source provenance and deterministic generation.
Start Isaac easily: double-click Start Isaac Sim.command, or run python3 scripts/isaac/launch.py. It opens a live Run Center and checks the GPU environment before marking it ready. Setup and cluster instructions.
Actual L40S recording on September 6, 2026: 16 / 16 selected demonstration cases crossed their openings. These are selected successes, not a random sample or the catalogue score. Full-resolution still · Recording and evaluation details.
Actual GPU run: September 6, 2026, 20:17:17–22:37:41 UTC, L40S, Isaac Sim 5.1.
| Coverage | Attempted | Audited traversals | Native errors |
|---|---|---|---|
| Complete non-pet collection | 985 / 985 | 44 / 985 (4.5%) | 10 |
| Upright doorway subset | 967 / 967 | 44 / 967 (4.6%) | 10 |
The unchanged Unitree checkpoint controls leg locomotion, with no handle or lock manipulation. This closed-start diagnostic is separate from the core benchmark. All 18 horizontal hatches were attempted but are not upright doorway tasks; errors remain in the denominator. Per-door results and limitations · Reproduce the run · Test your policy.
Dexterous humanoid development — one lever door:
Current status: the upright native robot can grasp the handle, partly open the door, transfer support to its left palm, return the lever and withdraw its right hand. Full qualified opening/traversal and learned vision/tactile control remain unfinished. Earlier complete native demonstrations predate the whole-handle audit and are not fully qualified under it.
| Engine / controller | Recorded result | UTC date |
|---|---|---|
| MuJoCo, upright privileged H1/Shadow teacher | 30/30 runtime checks + independent pad and whole-handle audits passed over 80.6 s; releases the handle with left-palm support at a 5.97° opening; no traversal (evidence) | September 9, 2026, 22:50 UTC verification |
| Isaac, upright privileged handle operation | Failed: 17/19 runtime checks over 36 s; latch operated, but fingertip contact and held opening were lost (evidence) | September 9, 2026, 23:54 UTC completion |
| Isaac, standing privileged grasp | 15/15 runtime checks, final opposed fingertip hold; contact accounting verified; no opening | September 9, 2026, 08:20 |
| Isaac, sensor-driven walking and stopping | 18/18 runtime + 22/22 independent checks over 10 s; no door interaction | September 9, 2026, 07:36 |
| MuJoCo, learned motor-target sensor actor | 0/1 complete tasks; fell after 1.306 s | September 9, 2026, 06:42 |
| MuJoCo, earlier privileged traversal with deep stance | Historical 47/47 task + 16/16 archive checks and same-start repeat; whole-handle qualification remains unverified | September 9, 2026, 02:34 |
Current plan and audit correction · Experiment history, failed trials and reproduction details.
Sensor locomotion evidence and limitations · Earlier failures, recovery curriculum and training commands.
The native sequence takes 127.99 seconds and uses privileged state; it is not a natural-speed motion or learned vision/tactile policy. Complete Isaac traversal, varied-start repeatability and catalogue coverage remain unfinished. Native evidence and replay instructions · Current progress · Isaac grasp evidence · Prepare a sensor demo.
Current plan and measured progress · Historical recordings, run times and discovered limitations.
The scripted-hand baseline was evaluated across all 985 benchmark doors, using three seeds for every assigned core scenario. It solved 849 / 985 doors (86.2%), with 3,727 / 4,182 successful episodes (89.1%). A solved door passed every scenario on every seed.
Why scripted attempts fail and when a test may be quarantined.
Explore every door’s result · Per-family and per-scenario tables · Open a recorded scripted attempt
These historical September 4, 2026 results describe the earlier geometry revision, with the 15 pet doors excluded. The scripted policy applies joint forces with a synthetic robot base. Recorded replays are separate, one-seed attempts; see their coverage and limitations.
This is a research dataset with known construction and physics approximations. Automated QA is not physical or artistic certification. Read the construction audit and appearance review before using it for training or evaluation.
The one-door physical human reference now opens, releases and walks through a passive door. All five digits maintain opposing contact throughout the working grasp; backward torso lean stays below 1.03°. Watch the full sequence with hand close-ups. This CPU simulation passes 18 physical checks. Its cautious gait remains a synthetic prototype, not captured human ground truth.
The human reference project aims to create natural simulated human demonstrations before retargeting them to robots. The earlier kinematic experiments do not meet the human-motion quality target and are not ground-truth human demonstrations.
git clone https://github.com/adamraudonis/DoorBench.git
cd DoorBench
python3 -m venv .venv
.venv/bin/python -m pip install -e ".[all]" huggingface_hub
.venv/bin/python scripts/huggingface_release.py download \
--revision v2026.09.05 --components assets --out data/doorbench
.venv/bin/python -m mujoco.viewer \
--mjcf data/doorbench/assets/doors/db0002_swing_single/scene.xmlUse DoorEnv for benchmark scenarios and access-control behavior; loading MJCF directly does not reproduce every environment callback. The release guide explains downloads, motion arrays, licensing and local catalogue setup. To regenerate data, see the dataset format and scripts/generate_dataset.py.
| Topic | Guide |
|---|---|
| Data layout and simulator use | Dataset format, integration, physics |
| Scenarios and historical baselines | Benchmark, results |
| Blender and vision data | Appearance, state bridge |
| Reference motion and review records | Motion schema and replay, human review |
| Humanoid experiments | Run your policy in Isaac Sim, MuJoCo G1 demo, Isaac Lab, current status |
Code, generated doors and original reference motion are MIT. Poly Haven textures retain CC0, including maps packed in Blender scenes. See release licensing and CITATION.cff; record the release tag and checksum inventory when citing experimental data.