CLARU
chocopanMIT

chocopan-t3-reverse-oracle-rlds-tipped-v5a

Synthetic scripted-oracle demonstrations of reverse manipulation tasks in simulation where objects are removed from containers and returned to the table, with realistic tipped object initializations. Dataset is provided in RLDS/TFDS format compatible with OpenVLA-OFT and LIBERO benchmarks.

Downloads13
Episodes3,477

Why This Matters for Physical AI

This dataset provides realistic object pose variation (upright and tipped) in container manipulation tasks, enabling training of manipulation policies that generalize to natural initial state distributions encountered in physical settings.

Technical Profile

Modalities
rgbproprioceptionlanguage
Robot Embodiments
Franka Panda
Action Space
end_effector_delta
Environment
simulation
Task Types
manipulationpick_and_placegrasping
Episodes
3,477
Data Format
RLDS
Annotation Types
language_instructionsreward_labels
License
MIT
Part of the LIBERO family

Access

View on HuggingFace

Need custom rgb data?

Claru builds purpose-built datasets for simulation applications with dense human annotations and quality assurance.

Request a Sample Pack

Related Datasets

TCoT LIBERO RLDS

RLDS/TFDS training data containing four filtered successful-demonstration suites and terminal windows of failed/retry rollouts used by libero_10_end60_retry experiments.

517 downloadsAug 2026other

ActionVAE processed training sidecars

Compact processed artifacts including paired LIBERO z-effect/z-action teacher caches, RoboTwin paired teacher caches, ActionAE sidecar protocols, and compressed LARYBench pilot archives for offline validation used in ActionVAE and phased LaWAM training pipelines.

86 downloadsSep 2026other

LIBERO-Plus_custom

Custom multi-view dataset composed of simulation scenes from LIBERO-plus assets, rendered from multiple viewpoints for finetuning the LVSM novel view synthesis model used in AnyCamVLA.

rgb
55 downloadsAug 2026MIT License (upstream LIBERO)

LIBERO GEAR Auxiliary Archive

A processed LIBERO-GEAR package retained as non-core historical/auxiliary data for the VLA SFT suite, containing tactile sensor data from robotic manipulation tasks.

tactile
52 downloadsJul 2026other

visual_robust_libero — Experiment 08: static embodiment supervision

A cross-embodiment robotic manipulation dataset built on LIBERO with 24 embodiments (6 robots × 4 grippers) designed to test whether supervising policies on static embodiment information improves task transfer. Includes action demonstrations, synthetic replay data, and visual question-answering annotations.

rgbdepthproprioceptionlanguage
45 downloadsSep 2026mit

ECOT Alignment 900 Episodes

A dataset of 900 rollout episodes generated by the MiniVLA policy across all 90 LIBERO-90 tasks, collected for supervised policy and reasoning alignment experiments. Both successful and failed episodes are included.

rgblanguageaction
40 downloadsSep 2026