zcyqyq/image_ccil

Image CCIL checkpoints

Selected DINO-transformer diffusion-policy and CCIL checkpoints.

Directory	Task / method	Checkpoint	Reported evaluation
square_baseline/	Square baseline	can_serl_diffusion_best.pt (epoch 500)	Offline: 80/150 = 53.33%, using 3 seeds and 50 episodes per seed (seed rates: 60%, 52%, 48%)
toolhang_baseline/	ToolHang baseline	can_serl_diffusion_best.pt (epoch 140)	Offline: 18/50 = 36.0%, using seed 42 and 50 episodes
square_e2e_be/	Square E2E, backward Euler (BE; legacy run)	can_serl_diffusion_epoch_360.pt	Inline training evaluation: 34/50 = 68.0%, using seed 42 and 50 episodes
square_e2e_zcurr_hybrid/	Square E2E, backward Euler from zcurr with hybrid latent space	can_serl_diffusion_epoch_500.pt	Offline: 104/150 = 69.33%, using 3 seeds and 50 episodes per seed (seed rates: 76%, 60%, 72%)
toolhang_e2e_be/	ToolHang E2E, backward Euler (BE)	can_serl_diffusion_epoch_240.pt	Offline: 12/50 = 24.0%, using seed 42 and 50 episodes total

The ToolHang E2E epoch-240 success rate above is a single-seed, 50-episode result; it is not a 3-seed average.

Each directory retains the corresponding run configuration, logs, metrics, and saved evaluation initial states. Only the selected checkpoint file is included from each run.

PushT

All PushT success rates below use the same offline protocol: diffusion inference seeds 42, 1042, and 2042; 50 test episodes per inference seed; environment seeds 100000..100049; and success threshold 0.95.

Directory	Method	Checkpoint	Offline success rate	Mean score
pusht/baseline/	Diffusion-policy baseline used to initialize the E2E and sequential experiments	policy.pt	57/150 = 38.00%	0.6113
pusht/e2e_best_epoch500/	End-to-end CCIL	policy.pt (epoch 500)	87/150 = 58.00%	0.7285
pusht/seq_be_epoch780/	Sequential CCIL, backward Euler	policy.pt (epoch 780)	77/150 = 51.33%	0.6697
pusht/seq_noisy_action_noise0_epoch960/	Sequential CCIL, noisy action with zero action noise and 0.5 quantile filter	policy.pt (epoch 960)	76/150 = 50.67%	0.6676

The two sequential policies share the standalone dynamics checkpoint and its training configuration in pusht/seq_shared_dynamics/. Their distinct data generation settings are stored as augmentation_config.yaml in the respective policy directories. The seq_noisy_action_noise0_epoch960 name describes the actual generator configuration; its historical local run directory was named pusht_seq_zcurr_q50.

Each PushT policy directory contains its exact configuration plus the 3-by-50 evaluation summary and per-inference-seed CSV. The baseline is approximately the commonly referenced 0.4-SR checkpoint; its larger standardized evaluation is reported precisely above as 0.38.

Square (structured artifacts)

The Square files below mirror the PushT layout. Standardized offline results use diffusion inference seeds 42, 1042, and 2042; 50 test episodes per inference seed; environment seeds 100000..100049; and success threshold 0.95.

Directory	Method	Checkpoint	Evaluation
square/baseline/	Diffusion-policy baseline	policy.pt (epoch 500)	80/150 = 53.33%
square/e2e_be_epoch360/	End-to-end CCIL, backward Euler (BE; legacy run)	policy.pt (epoch 360)	Inline seed 42: 34/50 = 68.00%
square_e2e_zcurr_hybrid/	End-to-end CCIL, backward Euler from zcurr with hybrid latent space	can_serl_diffusion_epoch_500.pt	104/150 = 69.33% (76%, 60%, 72%)
square/seq_be_epoch380/	Sequential CCIL, backward Euler	policy.pt (epoch 380)	92/150 = 61.33%
square/seq_noisy_action_noise0_epoch320/	Sequential CCIL, noisy-action generator with zero action noise and 0.5 quantile filter	policy.pt (epoch 320)	86/150 = 57.33%

The sequential policies share square/seq_shared_dynamics/dynamics.pkl; its exact training configuration, statistics, metrics, and the full extraction/training/ finetuning pipeline script are stored alongside it. Each sequential policy keeps its own augmentation configuration, finetuning metrics, and standardized 3-by-50 evaluation files. The legacy BE E2E epoch-360 number is a single-seed inline evaluation, not a standardized 3-by-50 result. The newer zcurr-hybrid E2E epoch-500 result is a standardized 3-by-50 offline evaluation (104/150).

Can (structured artifacts)

The Can files mirror the PushT and Square layouts. Standardized offline results use diffusion inference seeds 42, 1042, and 2042; 50 test episodes per inference seed; environment seeds 100000..100049; and success threshold 0.95. The baseline and sequential experiments use the first 20 demonstrations from the 40-trajectory (20%) source dataset, corresponding to 10% of the full dataset.

Directory	Method	Checkpoint	Offline success rate
can/baseline/	Diffusion-policy baseline used to initialize both sequential experiments	policy.pt (epoch 80)	59/150 = 39.33% (46%, 38%, 34%)
can/seq_be_epoch800/	Sequential CCIL, backward Euler	policy.pt (epoch 800)	78/150 = 52.00% (56%, 48%, 52%)
can/seq_noisy_action_noise0_epoch980/	Sequential CCIL, noisy action with zero action noise and 0.5 quantile filter	policy.pt (epoch 980)	73/150 = 48.67% (58%, 48%, 40%)

The two sequential policies share can/seq_shared_dynamics/dynamics.pkl and its training configuration. Their distinct generation settings are stored as augmentation_config.yaml in the respective policy directories. Each policy directory contains exactly one selected checkpoint, its 56 fixed Robomimic initial states, and the standardized 3-by-50 evaluation summary, JSON, and CSV. The complete extraction, dynamics, augmentation, and finetuning recipe is kept in can/run_can_seq.sh.