learner1119/ffw_sh5_n17_260820_left_h50_abs_30000

GR00T N1.7 · FFW-SH5 left arm · horizon 50 · absolute · step 30,000 / 50,000

Intermediate checkpoint of the run published as learner1119/ffw_sh5_n17_260820_left_h50_abs_50000 — same data, config and seed, saved at step 30,000 instead of the end. Use the final repo unless you specifically want the earlier point on the training curve.

action rep / tune_visual	ABSOLUTE / False
step	30,000 of 50,000
train loss at this step (25-pt moving avg)	0.0196
throughput	1.78 s/step on 4× A100 80GB
base	nvidia/GR00T-N1.7-3B (backbone nvidia/Cosmos-Reason2-2B, LLM layers ≤ 12)
global batch / lr / schedule	64 / 1e-4 / cosine, warmup 0.05, wd 1e-5, state_dropout 0.2

Train loss only — no validation split was held out. See the final repo's card for the full description of the data (left 8 of 16 dims; the right arm never moves), the modality config, and the four-way comparison.

Weights are BF16, 2 shards. Optimizer state (DeepSpeed ZeRO-2 shards) is not included. processor_config.json / statistics.json / embodiment_id.json sit at the repo root as the trainer wrote them. The modality config used for training is included as ffw_sh5_left8_h50_config.py; register it (import the file) before Gr00tPolicy(model_path="learner1119/ffw_sh5_n17_260820_left_h50_abs_30000", embodiment_tag="new_embodiment").

Siblings at the same step: ffw_sh5_n17_260820_left_h50_abs_vis_30000, ffw_sh5_n17_260820_left_h50_rel_30000.