File size: 3,303 Bytes
6b25ee0 ca6db9a 6b25ee0 ca6db9a 6b25ee0 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 | ---
library_name: pytorch
tags:
- robotics
- imitation-learning
- act
- diffusion-policy
- robotwin
license: mit
---
# Sigma_0: ACT-GAZER and DP-GAZER
24 student policies for six RoboTwin tasks, trained with **teacher visual KL = 0** and **student visual KL = 0.01**, training seed 0.
Code: [Cuixxx/Sigma_0 v1.0.0](https://github.com/Cuixxx/Sigma_0/tree/v1.0.0), commit `e02e6325a87738491b2ca182bdd244da47ab162c`.
| Family | Teacher prior | Tasks | Checkpoint |
|---|---|---:|---|
| ACT-GAZER | Gray-mask, all-camera | 6 | `policy_best.ckpt` |
| ACT-GAZER | Clean, all-camera | 6 | `policy_best.ckpt` |
| DP-GAZER-ATTN | Clean | 6 | `600.ckpt` |
| DP-GAZER-ATTN | Gray-mask | 6 | `600.ckpt` |
Tasks: `click_bell`, `place_empty_cup`, `place_container_plate`, `beat_block_hammer`, `lift_pot`, `stack_bowls_two`.
DP uses native 240 x 320 head-camera observations resized to 480 x 640 by the encoder. Both teacher and student stages ran for 600 epochs. ACT uses three cameras, 6000 configured training epochs, and the best validation checkpoint. ACT's action-latent KL is a separate setting from visual KL.
## Download
```python
from huggingface_hub import snapshot_download
snapshot_download(repo_id="Lucascui97/Sigma_0", local_dir="sigma0_weights")
```
To download one model, use `allow_patterns`, for example:
```python
snapshot_download(
repo_id="Lucascui97/Sigma_0",
local_dir="sigma0_weights",
allow_patterns=[
"DP-GAZER/gazer_distractor_strong-click_bell-aloha_agilex-joint-dp_gazer_attn-student-graymask-noteacherkl-res480-0/*",
"manifest.json",
],
)
```
## Loading and evaluation
Use the policy implementations and dependencies in the linked code release. These are custom policy checkpoints, not Transformers `from_pretrained` models.
- ACT: each experiment directory contains `policy_best.ckpt`, `dataset_stats.pkl`, and `training_config.json`, required by `XPolicyLab/policy/ACT_GAZER/model.py`. Configure `ckpt_dir` to that directory.
- DP: each experiment directory contains `600.ckpt`, the original final training payload with config, workspace states, optimizer and EMA states when present. The runtime adapter in `XPolicyLab/policy/DP/model.py` chooses the largest numeric checkpoint by default. `config.yaml` is a readable export of the checkpoint's embedded config.
- The DP experiment launchers also check for `latest.ckpt`. Before using them, create a local alias in each downloaded DP experiment directory:
```bash
for directory in sigma0_weights/DP-GAZER/*; do
if [ -f "$directory/600.ckpt" ] && [ ! -e "$directory/latest.ckpt" ]; then
ln -s 600.ckpt "$directory/latest.ckpt"
fi
done
export DP_GAZER_CHECKPOINT_ROOT="$PWD/sigma0_weights/DP-GAZER"
```
Training configs retain original provenance paths; update dataset and output paths for your environment. Teacher checkpoint files and training datasets are not included. Normal student inference uses RGB and robot state. Full training resumption may require the original data and teacher resources.
`manifest.json` lists each model, file size, and SHA-256 digest. Checkpoints have been checked against their stored training configuration; this upload does not represent a new simulation evaluation. Deployment on physical robots has not been validated by this release.
|