arkojit1/dit_pick_block_44000

Diffusion Policy on Ameyapores/pick_block_eef_delta, step 44,000

LeRobot Diffusion Policy (policy.type=diffusion, 278M parameters: ResNet18 encoder per camera + 1D conditional U-Net, DDPM with 100 timesteps), trained from ImageNet-initialised backbones on 31 of the dataset's 35 episodes (the last 4 held out).

• Inputs: observation.images.cam0, observation.images.cam2 (224x224), 4-dim observation.state, 2 observation frames
• Output: 4-dim end-effector delta action; horizon 64, 32 actions executed per inference
• Training: batch 64, Adam 1e-4 with cosine decay over a planned 100k steps, fp32; this is step 44,000 (~346 epochs)
• Held-out denoising loss at this step: 0.0722 (lowest seen: 0.0204 at step 5,000). That loss rises with overfitting but is a weak predictor of rollout success; this checkpoint has not been rolled out.

from lerobot.policies.diffusion.modeling_diffusion import DiffusionPolicy
policy = DiffusionPolicy.from_pretrained("arkojit1/dit_pick_block_44000")