ritamota/icub-vla-manipulation-t1-lerobot

iCub VLA Manipulation Dataset — T1 (LeRobot)

This repository contains Task 1 (T1) of the iCub VLA Manipulation Dataset, released in LeRobot format.

The dataset was developed to study the adaptation of generalist Vision-Language-Action (VLA) models, pretrained predominantly on conventional robotic manipulators, to manipulation with the humanoid iCub.

The complete dataset contains 500 successful simulated demonstrations across five tabletop manipulation tasks, with 100 demonstrations per task. Each task is available in both LeRobot and RLDS formats.

➡️ Complete dataset collection:
https://huggingface.co/collections/ritamota/icub-vla-manipulation-dataset

Task

T1 — Cube stacking

Pick up the red cube and put it on top of the green cube.

Each episode contains a successful execution of the task with randomised initial object positions within task-specific feasible regions.

Dataset Summary

Property	Value
Robot	iCub humanoid
Environment	Simulation
Task	T1 — Cube stacking
Episodes	100
Episode selection	Successful demonstrations only
Sampling frequency	5 Hz
Dataset format	LeRobot
Controlled end effector	Right hand
Action dimensions	7

Action Representation

Robot actions are represented by a common 7D Cartesian end-effector action vector:

[Δx, Δy, Δz, Δroll, Δpitch, Δyaw, g]

where:

• Δx, Δy, Δz are translational end-effector increments;
• Δroll, Δpitch, Δyaw are rotational increments;
• g represents the gripper command.

The Cartesian increments are interpreted in the simulation world frame.

Demonstration Generation

Demonstrations were generated in simulation using a task-specific scripted expert with access to privileged simulator state. This information was used only for demonstration generation and is not part of the information available to the VLA policies during training or evaluation.

The initial scene configuration is randomised between episodes to introduce intra-task variability while maintaining task feasibility. Only episodes satisfying the task-specific success criterion were retained.

LeRobot Format

The dataset is stored using the LeRobot dataset structure, including episode data and the associated metadata required for loading and processing the demonstrations.

The corresponding RLDS release of this task is available in the complete dataset collection linked above.

Related Tasks

The complete iCub VLA Manipulation Dataset contains:

Task	Description
T1	Pick up the red cube and place it on top of the green cube
T2	Move the bottle to the front of the laptop
T3	Push the green box to the centre of the table
T4	Remove the lid from the box
T5	Wipe the surface with the sponge

All five tasks are released separately in LeRobot and RLDS formats.

Intended Use

The dataset is intended for research on:

• Vision-Language-Action models;
• humanoid robot manipulation;
• imitation learning;
• embodiment-specific fine-tuning;
• cross-embodiment adaptation;
• robot-policy evaluation.

Limitations

• All demonstrations were generated in simulation.
• The dataset contains successful expert demonstrations only.
• Each repository contains a single manipulation task.
• The dataset was designed for embodiment-adaptation experiments with the iCub and should not be interpreted as a general-purpose humanoid manipulation benchmark.

Associated Work

This dataset was developed for the MSc thesis:

Application of VLA Models for Manipulation in Humanoid Robots
Rita Mota, Instituto Superior Técnico, Universidade de Lisboa, 2026.

Formal citation information will be added upon publication.

Licence

Except where otherwise noted, this dataset is released under the Creative Commons Attribution-ShareAlike 4.0 International (CC BY-SA 4.0) licence.

The dataset was generated using the pyCub simulator and includes rendered observations of the iCub humanoid robot. The official iCub model assets are distributed under CC BY-SA 4.0, while pyCub is distributed under CC BY 4.0. PyBullet, used as the underlying physics engine, is distributed under the zlib licence.

Third-party assets remain subject to their respective original licences and attribution requirements.