ChaoyiPan's picture
processed/chord: runs of the CPU workflow's current recipe
3d44b12 verified
|
Raw History Blame Contribute Delete
4.46 kB
# ARCTIC on two Sharpa hands: FlashCHORD scenes and SPIDER retargeting results
277 sequences of the [ARCTIC](https://arctic.is.tue.mpg.de) dataset (two hands manipulating an articulated
object) as self-contained MuJoCo scenes for two Sharpa Wave hands, each with the motion SPIDER's sampling MPC
retargets on it. They are the inputs and the reference outputs of SPIDER's CPU workflow (`examples/run_cpu.py`,
`docs/workflows/workflow-cpu.md`), whose scenes come from FlashCHORD (`docs/workflows/workflow-chord.md`).
## Layout
```
processed/chord/sharpa/bimanual/<task>/<data_id>/
scene.xml the MuJoCo model (meshes inline): two hands, the object, the table
scene.npz the reference motion and the tables of the control law
trajectory_cpu.npz the retargeted motion (qpos per 0.05 s frame), controls, planner statistics
metrics.json tracking errors and the pass / fail verdict of the run
visualization_cpu.mp4 the retargeted motion (solid) over the reference (green)
processed/chord/summary.csv one row per sequence: the names and the main numbers of metrics.json
```
`<task>` is `<subject>-<object>-<use|grab>` (`-retake` where ARCTIC has one) and `<data_id>` is ARCTIC's take number:
ARCTIC's `s01/box_use_01` is `s01-box-use/1`.
## Run it
```bash
git clone https://huggingface.co/datasets/retarget/retarget_example example_datasets
python examples/run_cpu.py # viewer: pick a task, press Start
python examples/run_cpu.py task=s01-laptop-use data_id=2 show_viewer=false num_threads=16
```
A run needs MuJoCo 3.17 or later (`mujoco.batch`), NumPy and Torch, and no GPU (SPIDER's README says how to set up
its environment); it replaces `trajectory_cpu.npz`, `metrics.json` and `visualization_cpu.mp4` of its scene
directory unless `output_dir` is set.
## What is here
- 277 sequences were retargeted with the workflow's default settings and seed 0; 267 of them pass
(object within 3 cm and 30 degrees of the reference on average, its hinge within 10 degrees, every opening
followed, the run not stopped). Means over all runs: object 1.13 cm and
3.4 degrees from the reference, wrists 0.90 cm, wrist shake
x1.39 of the demonstration's.
- A run that fails is here as it is: `summary.csv` and `metrics.json` say which and why.
- ARCTIC's 24 scissors sequences are not included: the articulated object models used here (ArtiGrasp) have
no scissors.
| object | sequences | pass |
|---|---|---|
| box | 25 | 25 |
| capsulemachine | 27 | 25 |
| espressomachine | 28 | 27 |
| ketchup | 28 | 27 |
| laptop | 29 | 28 |
| microwave | 27 | 27 |
| mixer | 29 | 29 |
| notebook | 25 | 24 |
| phone | 29 | 26 |
| waffleiron | 30 | 29 |
## How it was made
1. Reference: ARCTIC's raw sequences (MANO hand parameters and object poses) through the ARCTIC pipeline of NVIDIA's
video-to-data toolkit: MANO forward kinematics, inverse-kinematics retargeting to the Sharpa Wave hand, a support
surface under the object (widened to a table of 0.6 m radius). Articulated object models: ArtiGrasp's
`rsc/arctic`, with the hinge limit widened to the range ARCTIC demonstrates.
2. Scene: the model FlashCHORD compiles for that reference, exported with `python -m spider.simulators.chord_export`
(support fixed to the world, hinge held by friction, gains of the workflow's recipe baked in).
3. Retargeting: `python examples/run_cpu.py task=<task> data_id=<data_id> show_viewer=false seed=0`, MuJoCo 4.0.1,
with the defaults of `examples/config/cpu.yaml`: when the object slips the run goes back to before it was last
within 2 cm (1 s, then 2 s, then 3 s) and plans that stretch again looking 0.8 s ahead.
## Changes
- 2026-11-17: the runs were made again with the CPU workflow's current recipe (going back when the object slips,
the longer look ahead on the stretch planned again): 267 of 277 pass instead of 254. The files of
a run are now named after the CPU workflow (`trajectory_cpu.npz`, `visualization_cpu.mp4`); the scenes are
unchanged.
## Sources and terms
The motions and the object geometry derive from ARCTIC (Fan et al., CVPR 2023) and the object models from ArtiGrasp
(Zhang et al., 3DV 2024); the hand model is the Sharpa Wave hand as shipped with NVIDIA's toolkit. Their terms apply
to this derived data; ARCTIC is released for non-commercial research. No ARCTIC raw data and no MANO model files
are included.