Ada-4321/SkillWeaver-RLDS

SkillWeaver RLDS

Robot manipulation trajectories generated by SkillWeaver (CoRL 2026, arXiv:2609.36171), in RLDS (TFRecord) format, together with the LIBERO demonstrations they are co-trained with. This is the data used to fine-tune π0.5 for the LIBERO-Object tasks; the data processing and training code is in SkillWeaver-openpi.

Contents

libero_object/
├── train/task0 … task9/   # SkillWeaver trajectories, one directory per LIBERO-Object task
├── val/task0 … task9/     # held-out SkillWeaver trajectories
└── libero_demos/          # LIBERO-Object human demonstrations in the same format
samples/libero_object_task0/
├── trajectories/          # 4 raw SkillWeaver trajectories (NPZ) of LIBERO-Object task 0
├── scenes/                # their scene task specs
└── instruction_pool_manifest.jsonl   # their instruction pools

Split	Episodes	Steps
train	1,782	471,023
val	53	14,361
libero_demos	485	72,041

samples/ holds raw inputs of the conversion to RLDS. Converting them with SkillWeaver-openpi (examples/isaaclab_rlds/convert_npz_to_rlds.sh) reproduces the records of the same four episodes in libero_object/train/task0, apart from the source path in episode_metadata/file_path.

The LIBERO demonstrations are replayed from the official LIBERO-Object HDF5 files and converted to the same format; only successful replays are kept. Training uses train/ and libero_demos/; val/ is only used for validation metrics.

Format

Each directory holds TFRecord shards and a dataset_info.json with the feature dimensions and the episode and step counts. Each record is one episode. Arrays are stored as raw little-endian bytes (float32 unless noted).

Feature	Content
steps/observation/state	(T, 8): 7 arm joint positions and the gripper state
steps/action	(T, 8): absolute joint position targets and the gripper command
steps/action_gripper_binary	binary gripper command per step
steps/observation/exterior_image_{0..3}	JPEG frames, 256×256, from up to 4 exterior cameras; episode_metadata/exterior_view_count gives how many are present
steps/observation/wrist_image	JPEG frames from the wrist camera
episode_metadata/segment_{starts,ends,count}	sub-task segments of the episode
episode_metadata/segment_instruction_pools	language instructions for each segment (newline-separated)
episode_metadata/segment_instruction_pools_attr_aug	language-augmented instruction variants for each segment
episode_metadata/segment_hold_actions	action that holds the final pose of each segment
episode_metadata/camera_{ixt,ixt_hw,ext}_{0..3}, robot_base_pose_w	camera intrinsics, image size, extrinsics and robot base pose
episode_metadata/file_path	source trajectory of the episode

The parsing spec is examples/isaaclab_rlds/dataset_builder.py in SkillWeaver-openpi.

Download

pip install -U huggingface_hub
hf download Ada-4321/SkillWeaver-RLDS --repo-type dataset --local-dir SkillWeaver-RLDS

License

Apache-2.0. The LIBERO demonstrations are derived from LIBERO (MIT License).

Citation

@misc{zhu2026skillweaveragenticexplorationneural,
      title={SkillWeaver: Agentic Exploration over Neural Interaction Skills for Scalable Robot Data Generation},
      author={He Zhu and Lusen Zhao and Kwan Man Cheng and Su Li and Katerina Fragkiadaki},
      year={2026},
      eprint={2609.36171},
      archivePrefix={arXiv},
      primaryClass={cs.RO},
      url={https://arxiv.org/abs/2609.36171},
}