Request access to the FR3 native-10Hz Diffusion Policy checkpoints
This repository is publicly accessible, but you have to accept the conditions to access its files and content.
Access is reviewed manually by the repository owner. These checkpoints are provided for controlled robotics research and evaluation.
Log in or Sign Up to review the conditions and access this model content.
FR3 cube native-10Hz MimicGen Diffusion Policy
This repository contains evaluation-ready LeRobot Diffusion Policy checkpoints trained from scratch on the current native MimicGen 10Hz FR3 cube dataset. The repository is public and manually gated: each download request must be approved by the repository owner.
Immutable provenance
- Dataset:
jaeikkim/fr3-cube-full-episode-mimicgen-10k-rgb - Dataset revision:
87aac180ff7d659475bc903592d659649f4ca533 - Dataset size: 10,000 episodes / 1,416,273 frames
- Visual profiles: nominal lab 5,000 / lab variation 4,000 / stress tail 1,000
- Timebase: native MimicGen generation at 10Hz
- Input/output: three 320x180 RGB views + canonical 22D state -> 7D delta-EEF action
- Split: seed 42 episode-level 95/5; 1,345,458 train / 70,815 validation frames
- Training run:
AIDAS-Diffusion/fr3_cube_rgb_distillation/uglbqhxq - Hardware: 4 x A100 40GB, bf16
- Batch: 160/GPU, global 640
- Optimizer: Adam, LR
2.0726356840475473e-4, betas 0.95/0.999, weight decay1e-6 - Schedule: cosine with 1,200 update warmup
- Model: scratch ResNet18 + Conditional 1D U-Net; observation/prediction/execution = 2/16/8
- Completed target: 188,035 updates = 89.443446 train-split epochs
This run did not use the superseded ad056240... dataset lineage.
Released checkpoints
Training saved full-state checkpoints every 6,000 updates. The release selects the available saved checkpoint nearest to each requested 10-epoch target. A target epoch is never presented as the exact training epoch: directory names and this table retain the actual optimizer step.
| Target epoch | Actual epoch | Step | Model SHA256 |
|---|---|---|---|
| 10 | 11.416187 | 24,000 | 872d5ca92089002a3e3a304ed9bc9b2ae6ee82189acf001c436aa10ac15cfe67 |
| 20 | 19.978327 | 42,000 | b72bca92f4f2151e797184ce3bd2e65b9651b7ad8802c5aca694817943e98210 |
| 30 | 31.394514 | 66,000 | fbfbc0113dae33308b068af0278e494fab66df420c7154b4868a32be2a670b32 |
| 40 | 39.956654 | 84,000 | ca894d2cebcbe8ae2c65bbb87c34ebb09a62489b0c5bb4906a874732d4fa494d |
| 50 | 51.372841 | 108,000 | abb6fe898fc4a2e7d189dc57986c48ed85144bc9fdd43b16e8960527f8c0a958 |
| 60 | 59.934981 | 126,000 | 04225b473cfdeed33f031f9e6f45a5c8a8b14676f5a01380300aefb12758bf07 |
| 70 | 71.351168 | 150,000 | 3007d07d60ca826b07260bd0083cc3ed1f6f1a18e1a70e6a03fef454388ad0d6 |
| 80 | 79.913308 | 168,000 | 26ae3405780762c55f5d74f9147a7b1216e25ab4cd7fb5191ce9331cc167d9c8 |
| 90 | 89.443446 | 188,035 | 678130a805fef1888b5ba5062b70c2e3e97ccca613d51e0a576993224997f126 |
The uploaded folders contain pretrained_model inference assets only. Local
optimizer, scheduler, and RNG states are intentionally excluded.
Download and load
After the repository owner approves access, download one checkpoint with an authenticated Hugging Face account. For example:
from huggingface_hub import snapshot_download
from lerobot.policies.diffusion.modeling_diffusion import DiffusionPolicy
root = snapshot_download(
"jaeikkim/fr3-cube-mimicgen-native10hz-87aac180-diffusion-policy",
allow_patterns=["checkpoints/epoch-090-step-188035/*"],
)
policy = DiffusionPolicy.from_pretrained(
f"{root}/checkpoints/epoch-090-step-188035"
)
See release_manifest.json for the machine-readable provenance and mapping.