Abstract

In the field of robotic manipulation, deep imitation learning is recognized as a promising approach for acquiring manipulation skills. Additionally, learning from diverse robot datasets is considered a viable method to achieve versatility and adaptability. In such research, by learning various tasks, robots achieved generality across multiple objects. However, such multi-task robot datasets have mainly focused on single-arm tasks that are relatively imprecise, not addressing the fine-grained object manipulation that robots are expected to perform in the real world. This paper introduces a dataset of diverse object manipulations that includes dual-arm tasks and/or tasks requiring fine manipulation. To this end, we have generated dataset with 224k episodes (150 hours, 1,104 language instructions) which includes dual-arm fine tasks such as bowl-moving, pencil-case opening or banana-peeling,  and this data is publicly available. Additionally, this dataset includes visual attention signals as well as dual-action labels, a signal that separates actions into a robust reaching trajectory and precise interaction with objects, and language instructions to achieve robust and precise object manipulation. We applied the dataset to our Dual-Action and Attention (DAA), a model designed for fine-grained dual arm manipulation tasks and robust against covariate shifts. The model was tested with over 7k total trials in real robot manipulation tasks, demonstrating its capability in fine manipulation.

Paper Link

Autonomous Dual-Action Task Examples

Dual-Action and Attention Dataset

The Dual-Action and Attention (DAA) Dataset comprises 224,210 episodes of demonstrations, featuring dual-arm and/or fine manipulation skills, accompanied by gaze attention signals and dual-action labels that aid in learning fine manipulation. You can download the data from the following link (divided into each task group). The code for data conversion and data descriptions are available on our Github page.

Robot Configuration

Dual-Action and Attention

DAA leverages a gaze predictor to selectively focus on high-resolution pixels pertinent to the task, ensuring robust manipulation in the presence of unrelated objects or background noise. This system distinguishes between two types of actions: global-action, which generates a robust overall trajectory, and local-action, which allows for precise object manipulation. During both the demonstration phase and the model inference, the robot's trajectory is divided into these two components: the trajectory-based global-action for reaching movement and the reactive local-action for precise object manipulation. This dual-action mechanism ensures efficient and precise manipulation in dynamic environments.

We conducted an evaluation of our multi-task model's generalization ability across a variety of tasks and objects. This involved attempting 103 tasks with objects from the training set and 39 tasks with new objects not included in the training set. Task-specific models were trained for each object in each task group. The results reveal that the multi-task model achieved a 69.6% success rate across all tasks, significantly outperforming the task-specific models which only achieved a 12.2% success rate. Additionally, the multi-task model demonstrated effective generalization with a 61.5% success rate on tasks involving new objects.

Citation

If the DAA dataset is beneficial, please cite:

@article{kim2024multi,

  title={Multi-task robot data for dual-arm fine manipulation},

  author={Kim, Heecheol and Ohmura, Yoshiyuki and Kuniyoshi, Yasuo},

  journal={arXiv preprint arXiv:2401.07603},

  year={2024}}