dp-blockpick-dr-statenoise

LeRobot diffusion policy (ROBOTIS lerobot-cyclo fork @240b4a0, 287 M parameters) for a top-down block pick with the left arm of the ROBOTIS FFW SG2 in Isaac Lab (SG2 + ROBOTIS CAD gripper, calibrated wrist cameras, 29.25 in table, fixed yaml start pose, lime 5x5x4 cm block, grasp straight down with the fingertips 0.5 cm above the block's half height).

Files at the repo root = step 100k (DiffusionPolicy.from_pretrained("padiredd/dp-blockpick-dr-statenoise")); step-20000/, step-40000/, step-60000/, step-80000/ = intermediate checkpoints.

data	the domain-randomized 500-demo set (pick-cube-topdown-25cm-dr-500-3cam, 84,111 frames): per-episode random backdrop / table / lights / camera mounts / block colour AND varied demo timing; block uniform over x 0.23-0.48, y 0.065-0.315 m (the table edge is at x 0.27, so the demos span x 0.273-0.480)
cameras	head + both wrists, each stretched to 240x240, + 22-dim joint state; action = 16 joint targets, 15 Hz
image augmentation	heavy image augmentation (random order, up to 5 of: brightness 0.6-1.4, contrast 0.6-1.4, saturation 0.3-1.7, hue +-0.1, sharpness 0.3-2.0, affine +-10 deg / 10 % / 0.9-1.1, perspective 0.2, blur, random erasing, grayscale)
state noise	1 deg Gaussian noise on the 7 left-arm joint positions of the state input during training
recipe	100k steps, batch 8, Adam 1e-4 (0.95, 0.999), wd 1e-6, cosine, DDPM 100, horizon 64 / 32 executed, n_obs_steps 1
hardware	one ASU Sol A100 20 GB MIG slice, ~6 steps/s, ~4.5 h

Checkpoints uploaded: 20000, 40000, 60000, 80000, 100000. Sim evaluation results are added once they are run.