knightnemo/robocl-hdf-data-scaling-one-leg

HDF collection-budget scaling snapshot (one_leg, commit 1ee0338)

Resumable snapshot of the three-tier data-scaling experiment for the single-round heterogeneous-Diffusion-Forcing (HDF) world-model cycle on FurnitureBench one_leg (low randomness). Produced by Robo_Continual_Learning commit 1ee0338 (sweep configurations/sweep/furniture_hdf_data_scaling.yaml; see docs/furniture_hdf_data_scaling.md in that repo).

Tiers differ only in the frozen-base-policy collection budget; world-model gradient budget, fixed imagined-data budget, and replay capacities scale by the same factor f = episodes/1000.

State at snapshot (experiment stopped 2026-08-03)

tier	episodes	real transitions	WM steps done / target	last val NMSE
d1_1000	1,000	605,081	25K / 100K	0.0316
d2_4000	4,000	2,420,034	25K / 400K	0.0124
d3_16000	16,000	9,660,043	0 / 1.6M (collection done)	—

Frozen base policy evaluates at 50–52% success (4,096 rollouts) in all tiers.

Layout

• runs/hdf_<tier>_seed0/checkpoints/ — atomic resume sets (training_iteration_*.pt with WM dataset/optimizer/RNG/outer-loop state, fastsac_iteration_*.pt, real_replay_iteration_*.npz, replay_iteration_*.npz). The highest iteration index is the resume point.
• runs/hdf_<tier>_seed0/metrics.jsonl — per-iteration metrics including world_model/validation_normalized_mse and evaluation success rates.
• runs/hdf_<tier>_seed0/resolved_config.yaml, source_commit.txt.
• inputs/state_8000_stable_robocl.pt — frozen DPPO base policy (27 MB).
• inputs/world_model_best.pt — pretrained three-task H=100 HDF checkpoint, SHA-256 522a8e4990003808d3f7a1b60290674af18eaf921242d35b9701849418879cdf.
• inputs/one_leg_teleop_low_success.zarr/ — demonstration replay (2.1 GB).

Resuming

Point the training launcher at a checkout of commit 1ee0338, restore this snapshot to a local path, and set submission.resume_run_root (or experiment.resume_checkpoint) to the restored runs directory. The Python 3.8 runtime archive (Torch 2.0.1/CUDA 11.7 + IsaacGym + FurnitureBench) is NOT included because IsaacGym's license does not permit public redistribution; rebuild it with scripts/setup_furniture_dppo_env.sh.