VLA-90 PPO collection: released oracle_checkpoints are RL-keyed ('-v0'), and ENVS_CONFIG lacks the FindImposter*-VLA-v0 recipes

[sypsyp97]

Hi, thanks for releasing the VLA-90 task suite — the collection tooling is in great shape. While setting up self-collection for the -VLA-v0 PPO tasks I hit a gap I'd like to confirm, plus a question about the FindImposter family specifically.

What I observed (all on current main)

1. The published oracle checkpoints are keyed to the RL env ids, so they cannot drive any -VLA-v0 collection as-is. oracle_checkpoints.zip from avanturist/mikasa-robo contains 32 final_success_ckpt.pt, all under directories named like ShellGameTouch-v0 (the RL-32 set). mikasa_robo_suite/vla/dataset_collectors/get_mikasa_robo_datasets.py builds its checkpoint map from those directory names and matches --env-id exactly (no -VLA-v0 → -v0 aliasing), so e.g. --env-id ShellGameTouch-VLA-v0 raises Checkpoint for env_id=... not found.

2. No training recipe exists for the 9 FindImposter*-VLA-v0 tasks. vla/dataset_collectors/get_dataset_collectors_ckpt.py's ENVS_CONFIG has 68 entries (32 RL + 36 VLA, ending at index 67 = ShellGameColorLampTouch-VLA-v0); none of the FindImposter tasks appear. Meanwhile the collection side fully supports them (EPISODE_TIMEOUT_BY_ENV and parallel_dataset_collection_manager.py both list all 9), and the manifest marks them Data Source = PPO — so presumably you collected them from PPO oracles that aren't in the release.

3. The HF dataset repos avanturist/MIKASA-Robo-VLA-{npz,rlds,lerobot} currently contain only a README — I assume the VLA-90 data release is still in progress.

Questions

1. Do you plan to release -VLA-v0-keyed oracle checkpoints (ideally including the FindImposter family)? Any rough timeline for those and for the MIKASA-Robo-VLA-* data uploads?
2. If the checkpoints won't be released soon: could you share the PPO hyperparameters you used for FindImposter*-VLA-v0 (an ENVS_CONFIG-style entry would be perfect)? I'm happy to train the oracles myself and can PR the entries back.
3. Is reusing the RL -v0 checkpoints for VLA collection (by renaming the checkpoint directories) expected to work, given the VLA envs' randomized CUE_PHASE_STEPS / EMPTY_PHASE_STEPS? Or were the VLA datasets collected from oracles trained directly on the -VLA-v0 envs?

Thanks a lot!