PoMDAR
srl-ethz
A dexterous manipulation benchmark with simulation assets, printable objects and demonstration interfaces.
What you can explore
Explore OrcaHand environments and the documented webcam or Rokoko recording workflows.
Before you use it
Use the released task definitions and hardware configuration when comparing results. Printed object geometry and setup differences can change difficulty.
Original contributors
ETH Zürich SRL / PoMDAR contributors
License / access: Check original code, model and data terms before reuse
PoMDAR: inspect one dexterity task before comparing hands
- Environment
- Python 3.10/3.11 / MuJoCo viewer; optional teleoperation
- Embodiment
- Released Orca simulation model; not a generic hand driver
What you need
The top-level public README provides a minimal MuJoCo environment and a separate teleoperation environment. Begin with the passive viewer. A graphical display with working OpenGL is needed for that viewer.
First useful check
List the task IDs, inspect one scene and document the chosen task and asset revision before collecting performance data.
A small first experiment
Start with the minimal viewer dependencies
Use the root README environment, keeping optional webcam/glove packages out of the first check. Pin the checkout and record your Python and MuJoCo versions. The official repository remains the source for assets and their terms.
Public release and installation instructions →git clone https://github.com/srl-ethz/pomdar_benchmark cd pomdar_benchmark git rev-parse HEAD python3.10 -m venv .venv source .venv/bin/activate python -m pip install -r requirements.txt python -m pip checkUnderstand 18 tasks versus 17 files
H4 Palmar and H5 Pinch share H4_Palmar_H5_Pinch.xml. Count the benchmark procedures separately from the asset files. From sim/, use python launch_mujoco_orca.py --list-tasks to read the IDs accepted by your checkout.
Task selection implementation →Audit local task XML without opening the viewer
Run this original standard-library script from the repository root. It lists task IDs and catches malformed XML; it does not resolve includes, load meshes or validate physical dynamics. A missing folder usually means the wrong working directory or an incomplete checkout.
Task catalog →from pathlib import Path import xml.etree.ElementTree as ET folder = Path("sim/tasks") files = sorted(folder.glob("*.xml")) if not files: raise SystemExit("No task XML found; run from the repository root") for path in files: ET.parse(path) print(path.stem) print("Task asset files:", len(files)) print("H4 and H5 share one asset; file count is not benchmark task count") print("XML syntax only; meshes, physics and task success are not checked")Open one task, then vary one thing at a time
From sim/, run python launch_mujoco_orca.py --task V1_Wheel. Check the hand, wheel and scene before switching to H2_Chopsticks or G5_Cylinder. On macOS, MuJoCo passive-viewer scripts require mjpython in place of python. Save the task ID, asset revision and any parameter changes with a screenshot.
MuJoCo passive viewer and macOS launcher →Choose a teleoperation input deliberately
The webcam proof of concept moves fingers with a fixed hand base and does not provide reliable absolute wrist depth. The Rokoko example also controls wrist pose and calibrates from its first received pose. Neither path is automatically a scored, repeatable trial; check the upstream input and frame conventions first.
Webcam and Rokoko requirements →Define a comparison before collecting a score
Choose the same task procedure, objects, trial count, reset conditions and success rule for every hand. Report controller/operator, failures and any assistance. Use the paper and project protocol; do not infer a ranking from one attractive demonstration.
Design a comparable experiment →
What to record
- The selected task loads with its assets and the 17-file/18-task distinction is documented.
- Any claimed benchmark result has a task protocol and trial log beyond the viewer screenshot.
Common problems and what to check
- The task name is not found
- Use --list-tasks and an exact XML stem from the selected checkout; spaces in a paper name are not necessarily the CLI ID.
- Viewer fails on macOS or a headless server
- Use mjpython for the passive viewer on macOS and ensure a display/OpenGL is available. The XML inventory can run without the GUI.
- Webcam motion looks wrong
- Check occlusion and tracking quality; that example has a fixed hand base and does not recover full wrist pose.
- Asset reuse terms are assumed
- The root release notice and nested confidentiality notice conflict. Link to the source and obtain clarification before copying or redistributing the bundle.
Where it stops
Loading a scene is not completing a benchmark task. The public root README announces a release, but sim/README.md still carries a confidential-review notice. Reuse/distribution permission is therefore not resolved by this guide: confirm terms with the maintainers before redistributing assets. No source assets are republished here.
Keep exploring
ORCA Simulation
MuJoCo simulation environments for ORCA hands using the Gymnasium API.
ORCA Teleoperation
Capture-source and retargeting tools for operating ORCA hands from human hand motion.
ORCA Hand Hardware
CAD, printable parts and assembly resources for ORCA hand variants.
Editorial reference · checked 2026-10-24