transition-gen-world-policy-split

The same data as Gom-sy/transition-gen-world-policy, but with GT and generated episodes in separate dataset entries so that mix_ratio can control their proportion. Nothing was regenerated: the GT parquet files are byte-identical to the unsplit set (and therefore to the baseline Gom-sy/robocasa_coffee3_policy_groot_im128).

robocasa_panda_omron.{task}.gt    55 episodes   (original indices 0..54)
robocasa_panda_omron.{task}.gen   48 episodes   (renumbered 0..47; episode_manifest.jsonl
                                                 records source_episode_index)

dataset	ep	frames	usable samples (length − 15)
press_button.gt	55	6,561	5,736
press_button.gen	48	3,840	3,120
serve_mug.gt	55	17,752	16,927
serve_mug.gen	48	3,840	3,120
setup_mug.gt	55	15,342	14,517
setup_mug.gen	48	3,840	3,120

Why this split exists

In the first mixed run (47,977 steps, 30 epochs) the augmented policy scored about 5pp below the GT-only baseline, and only Press was significant (−6.8pp, p=0.001). The cause is not generated video quality — Press had the best transition reproduction of the three tasks (60%, against Serve 27%) and still dropped the most. The cause is that the generated share is uneven within each task, because GT episode lengths differ while 48 generated clips were added to each:

task	GT samples	generated samples	generated share	observed change
press_button	5,736	3,120	35.2 %	−6.8 pp
setup_mug	14,517	3,120	17.7 %	−3.3 pp
serve_mug	16,927	3,120	15.6 %	−4.6 pp

Sampling is frame-proportional, so GT under-exposure (−21% at matched steps) is identical across tasks and cannot explain the per-task spread. The generated share can.

mix_ratio settings

In the 6-entry form, mix_ratio is each entry's absolute sample share, not a multiplier on the dataset's size. Each entry holds one dataset, so relative_length is 1 and the weight reduces to mix_ratio; the mixture then draws dataset i with probability ∝ w_i / avg_shard_size_i and consumes a whole shard per draw, so the realised sample share is ∝ w_i and dataset size drops out.

That means setting every .gt entry to 1.0 would give each task a third of the GT samples, while the baseline weights GT by shard count (press 6, serve 17, setup 15 → 15.8 / 44.7 / 39.5 %). Press GT would be oversampled 2.11x and serve undersampled 0.74x — a second confound on top of the data change. Set the .gt entries to the baseline shard counts.

For a 15.6 % generated share within every task (multiplier 0.156 / 0.844 = 0.18483):

entry	mix_ratio		entry	mix_ratio
press_button.gt	6		press_button.gen	1.109
serve_mug.gt	17		serve_mug.gen	3.142
setup_mug.gt	15		setup_mug.gen	2.773

This gives 15.6 % generated in each task and overall, and leaves the GT task mix identical to the baseline. max_steps for matched GT exposure: 48,330 / 0.844 = 57,263.

For a 20 % share, use the multiplier 0.20 / 0.80 = 0.25 (press 1.5, serve 4.25, setup 3.75) and max_steps = 60,412. 15.6 % is the conservative choice: it is the share Serve already had, so no task's generated clips are drawn more often per epoch than they were in the first run.

Normalisation

None of the six sets ships a stats.json (relative_stats.json is empty), matching the baseline. Copy the baseline per-task stats.json into both the .gt and the .gen set of that task. If generate_stats is instead run over the generated sets, their values enter the min/max union and the size-weighted mean/std, which would move the mu-law eef bounds away from the baseline and add yet another variable.

Matched GT exposure still needs the extra steps

Controlling the share does not add data for free: at a fixed step count, raising the generated share lowers GT exposure. The baseline saw 48,330 GT-equivalent steps, so a "pure addition" arm needs 48,330 / (1 − share) steps.

Everything else — 128² storage, {0,1} gripper, control_mode/base_motion zero, 80-frame generated episodes, state_source: gt_window, and the two_labeler_agree flag (press 44/48, serve 27/48, setup 30/48) — is unchanged from Gom-sy/transition-gen-world-policy; see that card.