robopark's picture
Upload Pick Bucket BC and HandChord pretrained checkpoints
961e7c9 verified
|
Raw History Blame Contribute Delete
2.6 kB
---
library_name: pytorch
tags:
- robotics
- dexjoco
- flow-matching
- behavior-cloning
- hand-chord
---
# DexJoCo Pick Bucket pretrained checkpoints
Selected Pick Bucket BC policy and HandChord codec from DICE-RL. Both use seed 42 and the processed 98-demonstration dataset. The repository name preserves the requested spelling `dexjoco-pretrian-checkpoints`.
| Model | Checkpoint | Selection |
|---|---|---|
| Image BC flow policy | `pretrained_bc_policy_pick_bucket_img/checkpoint/state_2000.pt` | Epoch 2000 |
| HandChord | `pretrained_hand_chord_pick_bucket/checkpoint/codec.pt` | Final step 20000; latent dimension 8 |
Each model includes `.hydra/config.yaml` and the matching `normalization.npz`. `manifest.json` records source paths, file sizes, and SHA-256 checksums. Checkpoint bytes are unchanged. Packaged YAML files use repository-relative paths; original local training files are unchanged.
The policy uses front and wrist images (96 x 96), 23-dimensional proprioception, 22-dimensional actions, horizon 16, and 10 flow steps. The saved evaluation configuration uses a 700-step episode limit. The BC checkpoint retains its optimizer, scheduler, EMA, and RNG state. Its historical training-agent target is metadata; use the evaluation script below for inference.
## Download
From the DICE-RL repository root, after installing its DexJoCo dependencies:
```python
from huggingface_hub import snapshot_download
snapshot_download(
repo_id="robopark/dexjoco-pretrian-checkpoints",
repo_type="model",
local_dir="log_dir/dexjoco-pretrain",
)
```
## Evaluate the BC policy
```bash
MUJOCO_GL=egl python script/eval_pretrain_checkpoint.py \
--ckpt_path log_dir/dexjoco-pretrain/pretrained_bc_policy_pick_bucket_img/checkpoint/state_2000.pt \
--num_eval_episodes 10 --eval_n_envs 10
```
## Load HandChord
```python
from model.hand_chord.chunk_hand_chord import ChunkHandActionChord
chord = ChunkHandActionChord.load(
"log_dir/dexjoco-pretrain/pretrained_hand_chord_pick_bucket/checkpoint/codec.pt",
device="cpu",
)
```
HandChord uses the measured initial hand state, 16 hand joints, horizon 16, and 8 latent scalars per chunk. The retained training record reports validation MAE 1.7779712973625925 degrees at step 20000; this is a recorded metric, not a new evaluation.
Dataset: [robopark/dexjoco-single-arm-data](https://huggingface.co/datasets/robopark/dexjoco-single-arm-data).
## Upload from the staging folder
```bash
hf auth login --force
hf upload robopark/dexjoco-pretrian-checkpoints \
hf_release/dexjoco-pretrian-checkpoints --repo-type model
```