groot-blockpick-clean-run2-statenoise

GR00T N1.7 (3B) fine-tuned on the clean 500-demo top-down block-pick set (padiredd/pick-cube-topdown-25cm-500-3cam: ROBOTIS FFW SG2 in Isaac Lab, left-arm top-down pick of a lime-green block, 3 cameras at 240x240, 77,220 frames, fixed start pose, block uniform over x 0.23-0.48 / y 0.065-0.315 m, every demo with the same timing). Checkpoints live in step-<n>/ folders (this upload: step-20000, step-40000, step-60000, step-80000, step-100000); each folder loads with the ROBOTIS fork of Isaac-GR00T (omkarpatil18/Isaac-GR00T-n1.7 at e81d02b) via Gr00tPolicy(embodiment_tag="new_embodiment", model_path=<folder>) after registering the modality config modality_config/ffw_bg2_pick_cube_3cam_config.py (state = arm_left 8 + arm_right 8, action = the same 16 absolute joint targets, chunk 16, language "pick up the lime green block with the left arm").

Fine-tune recipe

base model	nvidia/GR00T-N1.7-3B (Cosmos-Reason2-2B backbone frozen; projector + diffusion action head tuned, 1.62 B trainable)
steps / batch	100,000 planned, global batch 32, checkpoint every 20k
optimizer	lr 1e-4 cosine, warmup 5 %, weight decay 1e-5, AdamW, fp32 weights
image aug	colour jitter brightness 0.4 / contrast 0.4 / saturation 0.5 / hue 0.1 + 5 deg random rotation + the repo's random crop
state	1 deg Gaussian noise on the 7 left-arm joints during training (our hook), state dropout 0.2 (repo default)
data sampling	shard 1024, episode sampling rate 0.1, 12 loader workers
hardware	ASU Sol A100 80 GB, ~0.68 s/step

Files per step folder: model shards (fp32, 12.6 GB), config.json, processor_config.json, statistics.json (normalization), embodiment_id.json, experiment_cfg/, trainer_state.json. Optimizer, scheduler and RNG states are not uploaded.

Evaluation

Sim rollouts (50 picks, same protocol as the diffusion policies in padiredd/dp-blockpick-*) are run with data_collection/rollout_policy.py --policy groot against groot_server.py; results are added here as they come in.