Should the adapter be trained on top of the teacher (PPO specialist) or the student (DAgger generalist)?

[benexyw-arch]

Hi, thanks for the great work!

I'd like to confirm the intended workflow for train_adapter. Per the paper, the pipeline reads as:
specialists → DAgger student (generalist) → adapter
So I expected the adapter to fine-tune on top of the DAgger generalist student. But the released code seems to only support loading from a PPO specialist teacher (logs/track/), not from a DAgger student (logs/dagger/).

The pretrained-checkpoint path is hardcoded to logs/track/ in train_adapter.py:
load_root = Path(WANDB_PATH_LOG) / "track" / args.load_exp_name / "checkpoints"

Even if I patch the path to logs/dagger/, the loader only accepts orbax-format numbered subdirs:
ckpts = [p for p in ckpt_root.glob("*") if p.is_dir() and p.name.isdigit()]

but train_dagger only saves PyTorch .pth files plus an ONNX export — no orbax checkpoint is ever written.

Could you clarify whether the adapter is intended to be fine-tuned from the teacher policy (PPO specialist) or the student policy (DAgger generalist)?Thanks!

[zzk273]

The adapter should be useful for both specialists and generalists. However, the current code indeed only supports loading pretrained checkpoints from specialists. I'll fix this soon.