mesh mesh |
|---|
Cup-on-hanger simulator
A standalone MuJoCo environment reconstructed for the cup-placement task, plus the original human demonstration used as its reference.
The task is to hang the blue cup by its handle on the upper front peg and the orange cup on the lower front peg, leaving both cups retained and the holder stable after release. The environment contains the two cups, hanger, tabletop, two YAM robot arms and three calibrated cameras.
Contents
examples/human_demo.mp4: original human example, 8.307 seconds, 1920×1080.model.xml: portable MuJoCo scene with an initializedstartkeyframe.scene/: scene-construction code and native/collision meshes.configs/: object/physics settings, workspace calibration and initial pose.simulate.py: viewer and headless simulation check.licenses/I2RT-LICENSE: license notice for the I2RT robot assets.
Setup and run
Download this private repository with Hugging Face authentication, then use Python 3.12 in a separate environment:
python3.12 -m venv .venv
source .venv/bin/activate
python -m pip install -r requirements.txt
python simulate.py
The viewer needs a desktop display and OpenGL. Drag in the viewer to inspect the scene. The robot arms hold their initial pose; this package provides the environment, not an autonomous cup-placement policy or a recorded agent run.
For a headless physics check:
python simulate.py --headless --seconds 3
To also render a PNG on a machine with EGL support:
MUJOCO_GL=egl python simulate.py --headless --seconds 3 --image preview.png
The model can also be loaded directly with mujoco.MjModel.from_xml_path.
Reset its start keyframe to restore the initialized robot/object state.
MuJoCo exposes the arm/gripper actuators and top_camera,
left_wrist_camera, and right_wrist_camera cameras. simulate.py shows the
original arm-hold/gravity-compensation convention.
Source and scope
This is the saved simulation from
tasks/cup_placement_cap/runs/cap_20261110_v1 in the local
Agentic Human2Sim2Real Benchmark. The MJCF geometry/physics and mesh files are
preserved; mesh references are made relative and a settled reset keyframe is
added. The scene builder uses bundled calibration instead of importing the
larger benchmark controller package. Its build() function can create a model
from the bundled scene configuration for further development.
The cups and hanger are reconstructed primitive geometry. This historical environment does not establish successful real-robot transfer or compliance with the later image-only policy-development protocol.
The demonstration matches the original attachment byte-for-byte:
cc844fcbd04808d279b0a8b3626ac5191d8cab8fb0bc1f6bf53ed5d36a1a46d3 (SHA-256).
- Downloads last month
- 31








