Square (robomimic) BC / CCIL checkpoints

Private release of Square PH absolute-action policies from the AI2 image_ccil Square grid (author helenw).

Companion PushT repo: helen9975/opal-pusht-ckpts.

Layout (by purpose)

Paths are family / data split / train mode — not dated folders.

{family}/data20|data_full/{expert_only,expert_aug}/best.pt
bc_{dino,resnet}/data20|data_full/best.pt
README.md  manifest.json

Families: bc_*, frozen_*, state_*, ft_*, seqa_*, seqb_*, seqc_*.

Quick download

huggingface-cli download helen9975/square-robomimic-ckpts \
  --include "state_tf/**" \
  --local-dir ./square_ckpts

# or whole repo
huggingface-cli download helen9975/square-robomimic-ckpts --local-dir ./square_ckpts

Each cell: best.pt + summary.json. Index: manifest.json.

Task & knobs

  • Env: robomimic Square PH, absolute EE actions, 2 cams
  • Image: 140 → crop [128,128]
  • Policy: diffusion; DINO patch_hw=[3,3] or ResNet-18 + UNet [256,512,1024]
  • Horizons: obs 1 / pred 16 / act 8; action_dim=10
  • CCIL eval: 3 seeds × 50 eps, success threshold 0.5
  • BC eval on card: in-loop 50 eps (indicative; not multi-seed)

Highlights

BC-only (in-loop 50):

Path recipe SR
bc_dino/data_full sb3 94.0%
bc_resnet/data_full sb4 82.0%
bc_dino/data20 sb1 60.0%
bc_resnet/data20 sb2 24.0%

Top EA — 20% demos (3×50):

Path recipe EA binary SR
state_unet/data20/expert_aug su1 70.7%
state_tf/data20/expert_aug st1 65.3%
ft_dino/data20/expert_aug d1 64.7%
seqa_dino/data20/expert_aug sa1 48.0%
seqb_resnet/data20/expert_aug s3 42.7%
seqb_dino/data20/expert_aug s1 34.7%
seqc_resnet/data20/expert_aug sc3 34.0%
ft_resnet/data20/expert_aug e1 28.7%

Top EA — full PH:

Path recipe EA binary SR
ft_dino/data_full/expert_aug d2 94.0%
state_tf/data_full/expert_aug st2 90.0%
seqa_dino/data_full/expert_aug sa2 90.0%
ft_resnet/data_full/expert_aug e2 89.3%
seqb_dino/data_full/expert_aug s2 87.3%
seqb_resnet/data_full/expert_aug s4 85.3%
state_unet/data_full/expert_aug su2 83.3%
seqc_resnet/data_full/expert_aug sc4 68.0%

Full leaderboard

Path recipe binary_SR note
bc_dino/data_full sb3 0.940 BC-only DINO full
ft_dino/data_full/expert_aug d2 0.940 FT BC→CCIL DINO full
ft_dino/data_full/expert_only d2 0.920 FT BC→CCIL DINO full
state_tf/data_full/expert_only st2 0.907 State TF→CCIL full
seqa_dino/data_full/expert_aug sa2 0.900 Seq-A DINO→CCIL full
state_tf/data_full/expert_aug st2 0.900 State TF→CCIL full
ft_resnet/data_full/expert_aug e2 0.893 FT BC→CCIL ResNet full (rerun wave)
ft_resnet/data_full/expert_only e2 0.893 FT BC→CCIL ResNet full (rerun wave)
seqb_dino/data_full/expert_aug s2 0.873 Seq-B DINO→CCIL full
seqb_dino/data_full/expert_only s2 0.860 Seq-B DINO→CCIL full
seqb_resnet/data_full/expert_aug s4 0.853 Seq-B ResNet→CCIL full
state_unet/data_full/expert_only su2 0.853 State UNet→CCIL full
seqa_dino/data_full/expert_only sa2 0.847 Seq-A DINO→CCIL full
state_unet/data_full/expert_aug su2 0.833 State UNet→CCIL full
bc_resnet/data_full sb4 0.820 BC-only ResNet full
seqb_resnet/data_full/expert_only s4 0.787 Seq-B ResNet→CCIL full
state_unet/data20/expert_aug su1 0.707 State UNet→CCIL
ft_dino/data20/expert_only d1 0.687 FT BC→CCIL DINO (corrected rerun)
seqc_resnet/data_full/expert_aug sc4 0.680 Seq-C ResNet→CCIL full
frozen_dino/data_full/expert_aug fd2 0.653 Frozen DINO→CCIL full
state_tf/data20/expert_aug st1 0.653 State TF→CCIL
ft_dino/data20/expert_aug d1 0.647 FT BC→CCIL DINO (corrected rerun)
seqc_resnet/data_full/expert_only sc4 0.633 Seq-C ResNet→CCIL full
bc_dino/data20 sb1 0.600 BC-only DINO 20%
state_unet/data20/expert_only su1 0.547 State UNet→CCIL
seqa_dino/data20/expert_aug sa1 0.480 Seq-A DINO→CCIL
seqb_resnet/data20/expert_aug s3 0.427 Seq-B ResNet→CCIL
frozen_dino/data_full/expert_only fd2 0.360 Frozen DINO→CCIL full
seqb_dino/data20/expert_aug s1 0.347 Seq-B DINO→CCIL
seqc_resnet/data20/expert_aug sc3 0.340 Seq-C ResNet→CCIL
state_tf/data20/expert_only st1 0.333 State TF→CCIL
seqb_dino/data20/expert_only s1 0.320 Seq-B DINO→CCIL
ft_resnet/data20/expert_aug e1 0.287 FT BC→CCIL ResNet (corrected rerun)
ft_resnet/data20/expert_only e1 0.287 FT BC→CCIL ResNet (corrected rerun)
frozen_resnet/data_full/expert_only fr2 0.260 Frozen ResNet→CCIL full
seqb_resnet/data20/expert_only s3 0.247 Seq-B ResNet→CCIL
bc_resnet/data20 sb2 0.240 BC-only ResNet 20%
frozen_resnet/data_full/expert_aug fr2 0.240 Frozen ResNet→CCIL full
seqa_dino/data20/expert_only sa1 0.227 Seq-A DINO→CCIL
frozen_dino/data20/expert_aug fd1 0.133 Frozen DINO→CCIL
seqc_resnet/data20/expert_only sc3 0.127 Seq-C ResNet→CCIL
seqc_dino/data_full/expert_aug sc2 0.120 Seq-C DINO→CCIL full
frozen_resnet/data20/expert_only fr1 0.107 Frozen ResNet→CCIL
frozen_resnet/data20/expert_aug fr1 0.087 Frozen ResNet→CCIL
seqc_dino/data_full/expert_only sc2 0.080 Seq-C DINO→CCIL full
frozen_dino/data20/expert_only fd1 0.027 Frozen DINO→CCIL
seqc_dino/data20/expert_aug sc1 0.020 Seq-C DINO→CCIL
seqa_resnet/data20/expert_only sa3 0.013 Seq-A ResNet→CCIL (corrected rerun)
seqc_dino/data20/expert_only sc1 0.013 Seq-C DINO→CCIL
seqa_resnet/data20/expert_aug sa3 0.000 Seq-A ResNet→CCIL (corrected rerun)

Recipe families

Family Codes Notes
BC-only sb1–sb4 DINO/ResNet × 20%/full
Frozen → CCIL fd* / fr* Frozen encoder, scratch head
State → CCIL st* / su* TF / UNet
FT BC → CCIL d* / e* Warm-start head from BC
Sequential → CCIL s* / sa* / sc* Seq-B / A / C SSL → freeze → CCIL
  • expert_only = latent FT on expert demos only (--expert_ratio 1.0)
  • expert_aug = expert + CCIL backward_euler_fast augmentations

Not in this release

  • seqa_resnet/data_full (no Seq-A ResNet full tab in Square grid)
  • obsolete dated HF prefixes square20_* / square_r1_* (removed earlier)
  • contaminated premature-resume dumps (archived under output/arxiv_premature_resume_20260722/)

Caveats

  • Contaminated premature-resume FT/sa3 weights were removed from this repo (old dated prefixes deleted). ft_* / seqa_resnet here are corrected reruns where applicable (d1/e1/e2/sa3 via rerun wave).
  • CCIL after strong BC is not guaranteed to improve Square (see e1: EA≈EO).
  • seqa_resnet 20% (sa3) remains near-zero after the corrected rerun.

Eval example

python eval_checkpoints_multi_seed.py \
  --ckpt ./state_tf/data_full/expert_only/best.pt \
  --output_dir ./eval_out --n_seeds 3 --n_test 50 --base_seed 42 \
  --success_threshold 0.5

Reproduce / re-stage

Local weka: image_embedding/image_ccil — see SQUARE_README.md.

python release/stage_square_ckpts_update.py
set -a && source scripts/ai2/secrets.env && set +a
export HF_HUB_OFFLINE=0 TRANSFORMERS_OFFLINE=0
unset HF_HUB_ENABLE_HF_TRANSFER
python release/upload_opal_to_hf.py \
  --repo helen9975/square-robomimic-ckpts \
  --release-dir release/square_ckpts_20260723 \
  --allow-overwrite
# then delete any leftover obsolete dated prefixes if needed
Downloads last month

-

Downloads are not tracked for this model. How to track
Video Preview
loading