CDIL datasets

Source- and target-domain datasets for Semi-Supervised Cross-Domain Imitation Learning (AdaptDICE). Paper: arXiv:2602.10793. Code: NYCU-RL-Bandits-Lab/CDIL.

Usage

The training script reads these files from ./dataset (--dataset_dir) by default. From the root of the code repo, run:

hf download rl-bandit-lab/CDIL --repo-type dataset --local-dir dataset

Then pass the filenames to train_il.py with --dataset_file_names="[<target expert 5>, <target expert 400>, <target random>]". See scripts/run_train.sh.

Files

Benchmark	Domain	Files
MuJoCo	Source (Hopper-v2, HalfCheetah-v2, Ant-v2)	{hopper,halfcheetah,ant}_{expert,random}-v2.hdf5 (unmodified D4RL, CC BY 4.0)
MuJoCo	Target (modified XMLs in env/)	target_{hopper,cheetah,ant}_{5,400}_*.npz (expert), target_{Hopper,HalfCheetah,Ant}-v3_random_{50,100,500}.npz
Robosuite	Source (Panda)	{Lift,Door,Wipe}_Panda_400_*.npz (expert), {Lift,Door,Wipe}_Panda_random_1600.npz
Robosuite	Target (UR5e)	{Lift,Door,Wipe}_UR5e_{5,400}_*.npz (expert), {Lift,Door,Wipe}_UR5e_random_{50,100,500}.npz

Expert files are named <env>_<num_trajs>_<avg_return>_<avg_episode_len>.npz. The 5-trajectory file supplies the labeled target expert demos. The 400-trajectory file supplies the expert part of the unlabeled imperfect data. The random_N file supplies the random part.

Each .npz contains init_obs, obs, action, next_obs, and done, with transitions stored trajectory after trajectory. Robosuite uses the JOINT_VELOCITY controller, object-state + robot0_proprio-state observations, reward_shaping=True, and horizon=500.

Env	Source state/action	Target state/action
Hopper	11 / 3	13 / 4
HalfCheetah	17 / 6	23 / 9
Ant	27 / 8	31 / 10
Lift	42 / 8	47 / 7
Door	46 / 8	51 / 7
Wipe	37 / 7	34 / 6

How the data was generated

• Expert data (everything except D4RL) was collected by rolling out SAC policies trained to expert level in each domain.
• Random data was collected by uniform-random rollouts. See utils.sample_demonstrations in the code repo (it collects fresh random rollouts when load_path=None).

Citation

@misc{chu2026semi,
  title         = {Semi-Supervised Cross-Domain Imitation Learning},
  author        = {Chu, Li-Min and Ma, Kai-Siang and Chen, Ming-Hong and Hsieh, Ping-Chun},
  year          = {2026},
  eprint        = {2602.10793},
  archivePrefix = {arXiv}
}