surgical-cosmos

Cosmos-Predict2 2B (480p) policy fine-tunes on SurRoL surgical tasks (cosmos-policy). Checkpoints are PyTorch distributed-checkpoint (.distcp) shards: checkpoints/iter_XXXXXXXX/{model,optim,scheduler,trainer}. model/ alone (3.7 GB) is enough for inference; optim/ (15 GB) is only for resuming training. Each run also carries its config.yaml, launch_info.yaml, job_env.yaml and wandb_id.txt.

Runs prefixed surrol_ come from surgical/cosmos-policy/outputs_*; runs prefixed cosmos_predict2_ come from cosmos-policy/output. Suffix _original = original Cosmos-Predict2 base, _surgical_h = Cosmos-H surgical base.

run	size
cosmos_predict2_2b_480p_knot_tying	19G
cosmos_predict2_2b_480p_needle_passing	19G
cosmos_predict2_2b_480p_surgical	55G
cosmos_predict2_2b_480p_surgical_raw	55G
surrol_bipegtransfer_10k_0to80_every10k_original	19G
surrol_bipegtransfer_10k_0to80_every10k_surgical_h	19G
surrol_gauzeretrieve_10k_0to80_every10k_original	19G
surrol_gauzeretrieve_10k_0to80_every10k_surgical_h	19G
surrol_gauzeretrieve_10k_retrain_0to80_original	19G
surrol_needlepick_10k_0to80_every10k_original	19G
surrol_needlepick_10k_0to80_every10k_surgical_h	19G
surrol_needlepick_10k_rand_0to80_every10k_original	19G
surrol_needlepick_10k_rand_0to80_every10k_surgical_h	19G
surrol_needlepick_aligned22d_grasp3_original	19G
surrol_needlepick_aligned22d_grasp3_surgical_h	19G
surrol_needlepick_compat_chunk8_grasp5_original	19G
surrol_needleregrasp_10k_0to80_every10k_original	19G
surrol_needleregrasp_10k_0to80_every10k_surgical_h	19G
surrol_pegtransfer_40k_rerun_20260725_original	19G
surrol_pegtransfer_40k_rerun_20260725_surgical_h	19G
surrol_pickplace_10k_oracle_chunk16	19G
surrol_pickplace_40k_oracle_chunk16	19G
surrol_pickplace_chunk16_long	19G
surrol_pickplace_chunk16_surgical_base	19G

Uploaded from /data/wenrui on 2026-09-08.