| --- |
| license: other |
| tags: [LeRobot, diffusion-policy, isaac-lab, simulation, ffw-sg2, robotics] |
| --- |
| # dp-blockpick-2k-withoutpush |
|
|
| LeRobot diffusion policy (ROBOTIS lerobot-cyclo fork, 288 M parameters) for a top-down block pick with the left arm of the ROBOTIS FFW SG2 (SG2 + ROBOTIS CAD gripper, calibrated wrist cameras, 29.25 in table). |
|
|
| Files at the repo root = step 100k (`DiffusionPolicy.from_pretrained("padiredd/dp-blockpick-2k-withoutpush")`); `step-10000/`, `step-20000/`, `step-30000/`, `step-40000/`, `step-50000/`, `step-60000/`, `step-70000/`, `step-80000/`, `step-90000/` = intermediate checkpoints |
| (every 10k steps). |
|
|
| | | | |
| |---|---| |
| | data | `pick-cube-2k-withoutpush-yaw45-full-dr2-3cam` (padiredd, on HF): 2,000 successful sim demos generated by the **code-as-policy program with the two CaP skill-layer defaults fixed** (data_collection/cap/programs/block_pick_v1_nocomp.py: close_gripper compensate=False, so the hand no longer rises 16 mm while closing, and --noise_scale 0, no executed-only joint pushes; one approach, close, lift; per-episode drawn pregrasp/lift/timing/via parameters). Same seeds (40001-40040, --parity_rng), block placements (|yaw| 0-45 deg), full region, table-height cycling and dr2 randomization as `padiredd/dp-blockpick-yaw45-2k` (motion planner) and `padiredd/dp-blockpick-cap2k-yaw45` (CaP with the defaults), so the three differ only by the demo generator. Diagnosis behind it: the cap2k DP placed 7/100 vs the planner's 64/100 because it copied the close-time lift (slips) and the pushes (hovers) | |
| | 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 A40 46 GB per run (local machine), ~5.7 steps/s (shared with sim data collection part of the time) | |
| |
| Task string: "pick up the lime green block with the left arm". Evaluation results are added once they are run. |
| |
