drawer_real60_bs128_ckpt

drawer_real60 diffusion checkpoints -- batch 128

Diffusion-policy behaviour cloning on LeeHakHo/drawer_real60 -- 60 real Franka demonstrations of the drawer task (open the drawer, pick the arrow, place it inside; two grasp cycles per demonstration), 30,374 samples at 15 Hz, 84px, seq_length 16, trained on one local RTX 4080 with the GPU to itself.

batch 128, lr 1.4e-4 (sqrt scaling of the 1e-4 used at batch 64), 700 warmup steps, cosine decay reaching 0 exactly at the final step, 1000 epochs x 100 steps -- the same recipe as LeeHakHo/square_real100_bs128_ckpt, so the tasks stay directly comparable. The run finished in 11 h 34 min.

ALL 60 demonstrations are training data -- hdf5_filter_key is null and experiment.validate is false -- so there is no validation curve and checkpoints can only be selected on training loss. That is what the reference diffusion-policy implementation does for its real-robot config. The file does ship mask/train 54 + mask/valid 6, deliberately unused here.

There is no simulator behind this dataset, so no success rate is reported here.

Arms

Baseline only, and that is a property of the dataset rather than a choice. obs/object and obs/aux_valid are absent from drawer_real60 -- the arrow's pose had not been tracked when the file was built -- so every auxiliary-pose arm fails at load time. A later dataset release adding the tracked arrow pose is what the aux arms are waiting on.

Training loss

Loss is the action loss.

Epochs

model_epoch_{100,200,...,1000}.pth -- the century checkpoints only. Each carries optimizer state and the EMA weights, so each is ~1.6 GiB.