dannymcy/libero_subtask_decomposed

CycleVLA Subtask-Decomposed LIBERO Dataset

This is the subtask-decomposed LIBERO RLDS dataset used to train CycleVLA, stored in RLDS (TensorFlow Datasets) format.

Paper: CycleVLA: Proactive Self-Correcting Vision-Language-Action Models via Subtask Backtracking and Minimum Bayes Risk Decoding

Project page: https://dannymcy.github.io/cyclevla/

Code: https://github.com/dannymcy/cyclevla_code

Dataset Description

CycleVLA decomposes long-horizon LIBERO manipulation tasks into subtask segments, each labeled with a subtask language instruction and per-step progress values. This dataset combines all four LIBERO suites (Spatial, Object, Goal, and LIBERO-10) into a single set of subtask sub-episodes, with extended supervision signals for training self-correcting VLA policies.

Each full demonstration is split into subtask chunks using a VLM-based decomposition pipeline (see the paper for details). Terminal frames are oversampled following a NaVILA-style strategy:

• Gripper subtasks: last frame repeated 8 times
• Non-gripper subtasks: last 3 frames repeated 4 times each

Dataset Statistics

Metric	Value
Sub-episodes	6,501
Total transitions	266,207
LIBERO suites	Spatial, Object, Goal, LIBERO-10
TFRecord shards	128

Features

Feature	Shape	Type	Description
observation/image	(256, 256, 3)	uint8	Third-person camera RGB
observation/wrist_image	(256, 256, 3)	uint8	Wrist camera RGB
observation/state	(8,)	float32	Robot EEF state (6D pose + 2D gripper)
observation/joint_state	(7,)	float32	Robot joint angles
action	(7,)	float32	Robot EEF action (6D delta + gripper)
language_instruction	—	string	Subtask language instruction
is_first	—	bool	First step of the sub-episode
is_last	—	bool	Last step (stop signal s_t)
is_terminal	—	float32	Subtask progress p_t ∈ [0.1, 1.0]
reward	—	float32	1.0 on final step, 0.0 otherwise

9-Dimensional Action Space (as used in training)

During training, the raw features are assembled into a 9-dim action vector:

a_t = [Δx, Δy, Δz, Δu, Δv, Δw, γ, s_t, p_t]
       ├─── 6D EEF delta ───┤  grip  stop  progress

• Dims 0–5: End-effector pose deltas (from action[:6])
• Dim 6 (γ): Gripper action (from action[6])
• Dim 7 (s_t): Binary stop signal (from is_last)
• Dim 8 (p_t): Subtask progress in {0.1, 0.2, ..., 0.9, 1.0} (from is_terminal)

Loading the Dataset

With TensorFlow Datasets

import tensorflow_datasets as tfds

# Download from HF first, then load from the local directory
builder = tfds.builder_from_directory("path/to/libero_subtask_decomposed")
ds = builder.as_dataset(split="train")

for episode in ds.take(1):
    for step in episode["steps"]:
        image = step["observation"]["image"]
        action = step["action"]
        language = step["language_instruction"]
        progress = step["is_terminal"]  # p_t ∈ [0.1, 1.0]
        stop = step["is_last"]          # s_t ∈ {0, 1}

Download

# Install huggingface_hub
pip install huggingface_hub

# Download all files
huggingface-cli download dannymcy/libero_subtask_decomposed --repo-type dataset --local-dir ./libero_subtask_decomposed

Citation

@article{ma2026cyclevla,
  title={CycleVLA: Proactive Self-Correcting Vision-Language-Action Models via Subtask Backtracking and Minimum Bayes Risk Decoding},
  author={Ma, Chenyang and Lu, Kai and Yang, Guangyu and Liu, Jiuming and Xu, Shitong and Byrne, Bill and Havoutis, Ioannis and Trigoni, Niki and Markham, Andrew},
  journal={arXiv preprint arXiv:2601.02295},
  year={2026}
}

License

Apache 2.0