Robot Learning from Demonstration: Enhancing Plan Execution with Failure Detection Model

Tianyi Wu, Liwei Shen, Zhen Dong, Xin Peng, Wenyun Zhao

Primary Area: applications to robotics, autonomy, planning

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Keywords: action failure detection, learning from demonstration, meta learning

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.

Abstract:

Learning plans from demonstrations has emerged as a valuable paradigm, in which a robot autonomously completes a task by executing a sequence of actions according to a learned plan.
Nevertheless, the execution of an action may encounter failures in the real environment, such as failing to pick up a cup, resulting in plan execution failure.
The execution of a broken plan may damage the environment, e.g., cooking coffee when a cup is not successfully placed.
To avoid such risks, action failure detection is crucial.
However, the action failure within the execution of task plans is often neglected in existing research.
To address the problem, we propose a framework that learns an executable plan that checks failures of each action, called failure-aware plan.
Our framework employs meta-learning to learn neural network-based failure-aware task plans.
Initially, by using trajectory data collected from robot randomness execution, the framework pre-trains a model that discriminatively captures the state features of various actions at different stages.
Utilizing user demonstration trajectories labeled as either success or failure, the pre-trained model undergoes fine-tuning, which is then employed to determine the success or failure of an action execution by means of the corresponding state features.
We demonstrate the effectiveness of our approach through experiments on a robot in a simulation environment.
Our approach outperforms the compared method when only limited demonstration data is available.
This work contributes to enhancing the reliability of plan execution for robot by considering action failure detection.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors' identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 7916