We release the plot data (ogbench-agg.pkl, ogbench-individual.pkl, robomimic-individual.pkl) to facilitate future research. Each file is a dictionary keyed by (task_name, method_name) tuples. Each entry contains a nested dictionary with four arrays:
steps: training steps (in millions),means: mean success rate at each step,ci_lows,ci_highs: lower and upper bounds of the 95% bootstrap confidence interval with 5000 samples. By default, all arrays contain 21 values (recorded every 100K steps from 0 to 2M). For online baselines (RLPD, RLPD-AC, QC-RLPD, SUPE-GT), the arrays contain only 11 values instead.
