Dataset card: the SO101 pick-and-place study (Part E, folder so101_pick_place/)
Browse files
README.md
CHANGED
|
@@ -59,14 +59,16 @@ configs:
|
|
| 59 |
data_files: raw/tracked/episodes.parquet
|
| 60 |
---
|
| 61 |
|
| 62 |
-
# NeDM datasets: the paper, the traversing study, the contact NRD
|
| 63 |
|
| 64 |
This repository holds separate releases. **Part A** is the set of datasets behind the NeDM paper, unchanged since its
|
| 65 |
release (Hub tag `paper-v1`). **Part B**, everything under `traversing/`, holds the drives, training files, models and
|
| 66 |
evaluation records of the traversing study, follow-on work that is not part of the paper. **Part C**, everything under
|
| 67 |
`contact_nrd/`, holds the training data, test sets and models of the contact NRD, also follow-on work. **Part D**,
|
| 68 |
everything under `g1_footstep_rl/`, holds the policies, RL inputs and MuJoCo results of a G1 footstep policy trained in
|
| 69 |
-
the contact NRD, also follow-on work.
|
|
|
|
|
|
|
| 70 |
|
| 71 |
Contents:
|
| 72 |
|
|
@@ -75,6 +77,7 @@ Contents:
|
|
| 75 |
- [Part B: traversing study](#part-b-traversing-study): `traversing/`
|
| 76 |
- [Part C: contact NRD](#part-c-contact-nrd): `contact_nrd/`
|
| 77 |
- [Part D: G1 footstep policy](#part-d-g1-footstep-policy): `g1_footstep_rl/`
|
|
|
|
| 78 |
|
| 79 |
## Part A: datasets of the paper
|
| 80 |
|
|
@@ -793,3 +796,71 @@ pinned upstream commit (OpenHomie, commit `33071f79`), which has its own license
|
|
| 793 |
BSD-3-Clause, the same as Part A and the code. HOMIE is not hosted here; it is fetched from its pinned upstream commit,
|
| 794 |
under its own license. The G1 footstep policy has no publication yet: please cite the NeDM paper (Part A) and name the
|
| 795 |
dataset revision you used (`cba82611210e401dc0155cb78fcf456a385ecef4`).
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 59 |
data_files: raw/tracked/episodes.parquet
|
| 60 |
---
|
| 61 |
|
| 62 |
+
# NeDM datasets: the paper, the traversing study, the contact NRD, the G1 footstep policy and the SO101 pick and place
|
| 63 |
|
| 64 |
This repository holds separate releases. **Part A** is the set of datasets behind the NeDM paper, unchanged since its
|
| 65 |
release (Hub tag `paper-v1`). **Part B**, everything under `traversing/`, holds the drives, training files, models and
|
| 66 |
evaluation records of the traversing study, follow-on work that is not part of the paper. **Part C**, everything under
|
| 67 |
`contact_nrd/`, holds the training data, test sets and models of the contact NRD, also follow-on work. **Part D**,
|
| 68 |
everything under `g1_footstep_rl/`, holds the policies, RL inputs and MuJoCo results of a G1 footstep policy trained in
|
| 69 |
+
the contact NRD, also follow-on work. **Part E**, everything under `so101_pick_place/`, holds the NRD, data,
|
| 70 |
+
checkpoints and real-robot records of an SO101 pick-and-place camera policy trained only in a contact NRD, also
|
| 71 |
+
follow-on work.
|
| 72 |
|
| 73 |
Contents:
|
| 74 |
|
|
|
|
| 77 |
- [Part B: traversing study](#part-b-traversing-study): `traversing/`
|
| 78 |
- [Part C: contact NRD](#part-c-contact-nrd): `contact_nrd/`
|
| 79 |
- [Part D: G1 footstep policy](#part-d-g1-footstep-policy): `g1_footstep_rl/`
|
| 80 |
+
- [Part E: SO101 pick and place](#part-e-so101-pick-and-place): `so101_pick_place/`
|
| 81 |
|
| 82 |
## Part A: datasets of the paper
|
| 83 |
|
|
|
|
| 796 |
BSD-3-Clause, the same as Part A and the code. HOMIE is not hosted here; it is fetched from its pinned upstream commit,
|
| 797 |
under its own license. The G1 footstep policy has no publication yet: please cite the NeDM paper (Part A) and name the
|
| 798 |
dataset revision you used (`cba82611210e401dc0155cb78fcf456a385ecef4`).
|
| 799 |
+
|
| 800 |
+
## Part E: SO101 pick and place
|
| 801 |
+
|
| 802 |
+
### What it is
|
| 803 |
+
|
| 804 |
+
Follow-on work to the paper, not part of it. A camera policy (the "student") for a real SO101 arm picks a 30 mm cube
|
| 805 |
+
and places it on a plate. Every training step ran inside a contact NRD of the arm and the cube (the design of Part C,
|
| 806 |
+
trained on CPU MuJoCo 3.15.0 episodes): a state teacher trained with PPO in the NRD, then the student trained in the
|
| 807 |
+
NRD with ray-traced images of a calibrated copy of the real scene (online DAgger). No real image was a training
|
| 808 |
+
sample. On the real arm the student placed the cube in 36 of 50 stress-test episodes (72 %; 95 % interval 58–83 %).
|
| 809 |
+
The training, the results and the commands are in
|
| 810 |
+
[`so101_pick_place/` of the code repository](https://github.com/uwsbel/NeDM/tree/main/so101_pick_place).
|
| 811 |
+
|
| 812 |
+
### Layout
|
| 813 |
+
|
| 814 |
+
| Path | Contents |
|
| 815 |
+
|---|---|
|
| 816 |
+
| `so101_pick_place/nrd/c1_s62/`, `nrd/c0_s62/` | The policy NRD (every teacher and student stage ran in it) and its base: `best.pt` loads with `nedm.contact_nrd.model.load_model`; `original/best.pt` is the file as trained (two training-only heads more); the training records |
|
| 817 |
+
| `so101_pick_place/nrd_data/` | The 4 NRD training data parts (93,494 episodes; `unified_data.npz`, `system.json`, `manifest.json`), and the 7 PPO policies behind `policy_v3` |
|
| 818 |
+
| `so101_pick_place/nrd_test/` | The sealed scripted test set (3,050 episodes) |
|
| 819 |
+
| `so101_pick_place/teacher/`, `student/` | Every teacher (T0-T6) and student (S0-S5) checkpoint of the chain with its run record; the teacher's expert data |
|
| 820 |
+
| `so101_pick_place/rl_inputs/` | Task sets, start banks, geometry, the task sets of the sim checks |
|
| 821 |
+
| `so101_pick_place/results/` | The sealed NRD test, the sim-check results and selection rules, the per-episode robot results |
|
| 822 |
+
| `so101_pick_place/robot/` | Raw logs of 19 robot sessions (102 policy episodes), 52 videos of the student's input images, 2 LeRobot teleop recordings |
|
| 823 |
+
| `so101_pick_place/bundle/` | The folder installed on the robot computer, and its tarball |
|
| 824 |
+
| `so101_pick_place/scene_calib/`, `provenance/` | Camera calibration, scene and servo fit (also in the code repository); code snapshots that are in no commit |
|
| 825 |
+
| `so101_pick_place/release_manifest.json` | Size and SHA-256 of every file; provenance (commits, runs, sources) |
|
| 826 |
+
|
| 827 |
+
Total 13.3 GB (NRD data 10.6 GB).
|
| 828 |
+
|
| 829 |
+
### What the files contain
|
| 830 |
+
|
| 831 |
+
- NRD data (`unified_data.npz`, uncompressed): `arm` [N, 601, 12] (6 joint angles, 6 joint rates), `cube` [N, 601,
|
| 832 |
+
13] (position, quaternion wxyz, velocity, angular velocity), `action` [N, 601, 6] (joint position targets),
|
| 833 |
+
`contacts` [N, 600, 3] (fixed jaw - cube, moving jaw - cube, cube - table), `lengths`, `splits` (0 train,
|
| 834 |
+
1 validation, 2 test), `family`, `phase`, contact force, impulse and depth, ids, seeds and outcome fields. Records
|
| 835 |
+
are 10 ms apart, episodes 4 s.
|
| 836 |
+
- Checkpoints: PyTorch files. Teachers: MLP state policies with privileged inputs. Students: two ResNet18 image
|
| 837 |
+
encoders (external camera crop and wrist camera, 224 x 224 RGB), joint angles and rates, a chunk of 10 joint
|
| 838 |
+
targets as output.
|
| 839 |
+
- Robot logs: per session `session.json` (command line, typed outcome per episode) and `first_frames.npz`; per
|
| 840 |
+
episode `episode_live.json` (setup, guard, latency, outcome) and `episode_live.npz` (50 Hz joints and commands,
|
| 841 |
+
each policy query with its action chunk and the two input images at 112 px).
|
| 842 |
+
|
| 843 |
+
### Download and use
|
| 844 |
+
|
| 845 |
+
```bash
|
| 846 |
+
scripts/so101_pick_place/reproduce.sh full --steps download # in the code repository; checks every SHA-256
|
| 847 |
+
```
|
| 848 |
+
|
| 849 |
+
The code repository's `so101_pick_place/release_manifest.json` pins the Hub revision; `--with-nrd-data` and
|
| 850 |
+
`--with-robot` add the large folders. The MuJoCo Menagerie SO101 model is not in this dataset: the code repository
|
| 851 |
+
fetches it at a pinned revision.
|
| 852 |
+
|
| 853 |
+
### Known limitations
|
| 854 |
+
|
| 855 |
+
- **One robot, one lab, one cube and plate.** The outcome is typed by the operator; the stress protocol was not fixed
|
| 856 |
+
in advance, and the student was chosen with robot results.
|
| 857 |
+
- **The NRD test set has only scripted families** and is not fully blind (run once before on an older NRD design).
|
| 858 |
+
- **Lab content.** The robot logs hold lab background images, the robot computer's user path and the USB adapter
|
| 859 |
+
serial. The videos show the student's input images as logged on the robot (112 px; the student sees 224 px);
|
| 860 |
+
there is no full-resolution video.
|
| 861 |
+
|
| 862 |
+
### License and citation
|
| 863 |
+
|
| 864 |
+
BSD-3-Clause, the same as Part A and the code. The MuJoCo Menagerie model is not hosted here. The SO101 pick and place
|
| 865 |
+
has no publication yet: please cite the NeDM paper (Part A) and name the dataset revision you used (the
|
| 866 |
+
`hf_revision` of `so101_pick_place/release_manifest.json` in the code repository).
|