You need to agree to share your contact information to access this model

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

FR3 Cube MimicGen10K Diffusion Policy

Evaluation-ready LeRobot Diffusion Policy checkpoints at cumulative optimizer steps 10K, 20K, 30K, 40K, 50K.

Policy contract

3 native 320x180 RGB views
observation horizon: 2
prediction horizon: 16
action execution chunk: 8
ResNet18 visual encoder
canonical 22D state
7D delta EEF/OSC action (relative Cartesian 6D + binary gripper)

The source dataset stores a 23D xyz+quaternion state. The runtime adapter maps it to the same canonical 22D xyz+rotation-vector state used by the OmniReset comparison. The action space is unchanged.

Dataset and training

Dataset: jaeikkim/fr3-cube-full-episode-mimicgen-10k-rgb
W&B: yis55z5q
2 x NVIDIA B200, batch/GPU 768, global batch 1536
base LR: 4.974325641714113e-4, cosine scheduler, 500-step warmup

These exports contain the policy, processors, and normalization statistics for evaluation. Optimizer/scheduler state is intentionally not published, so they are not full-state training-resume checkpoints.

Download

hf download jaeikkim/fr3-cube-mimicgen10k-diffusion-policy \
  checkpoints/step_050000/pretrained_model \
  --local-dir fr3_cube_mimicgen_dp

Use checkpoint_manifest.json and SHA256SUMS to verify the release.

Downloads last month: -; Downloads are not tracked for this model. How to track

Video Preview

Robotics