dp-openbottle-v4-visual-dr-200

LeRobot diffusion policy (ROBOTIS lerobot-cyclo fork, 288 M parameters) for opening a bottle with both arms of the ROBOTIS FFW SG2 (left holds the bottle, right unscrews the cap).

Files at the repo root = step 100k (DiffusionPolicy.from_pretrained("padiredd/dp-openbottle-v4-visual-dr-200")); step-10000/, step-20000/, step-30000/, step-40000/, step-50000/, step-60000/, step-70000/, step-80000/, step-90000/ = intermediate checkpoints (every 10k steps).

data	open-bottle-v4-visual-dr-200-3cam (not on HF): the first 200 successful sim demos of the open-bottle v4 task (data_collection/bottle_cap/collect_open_v4.py, 2026-11-08): the left arm picks a 28 oz bottle by its lower end and holds it at a station (lean 0-15 deg), the right hand unscrews the cap in up to 3 x 50 deg strokes (release 140-149 deg), the cap stays in the hand / pops off / falls, and the right hand ends holding the cap; --dr_visual (table look, backdrop, lights, spot lights); only successful episodes, 3 cams stretched to 240x240 (snapshot taken while the 1,000-demo collection continued)
sampling	uniform over all frames (LeRobot default EpisodeAwareSampler)
inputs / outputs	head + left wrist + right wrist cameras, each 240x240, + 22-dim joint state; action = 16 joint targets, 15 Hz
image augmentation	heavy, 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	none
recipe	100k steps, batch 32, ResNet18 per camera, Adam 1e-4, cosine, DDPM 100 steps, horizon 64 / 32 executed, n_obs_steps 1, seed 1000
hardware	one NVIDIA RTX 4090 24 GB (lab225), ~10.4 steps/s, ~2 h 40 min

Task string: "open the bottle: hold it by the base with the left hand, unscrew the cap with the right hand". Evaluation results are added once they are run.