The paper addresses the challenge of automating long-horizon robotic tasks, which require the execution of multiple actions to complete a task. Optimization-based planning methods, such as task and motion planning (TAMP), are efficient for creating action plans for long-horizon tasks. However, these methods require manually designed conditions to avoid unreasonable states, such as collisions between objects. The design of these conditions has two critical issues: 1) it is time-consuming due to the trial-and-error process, and 2) it is difficult to manually cover all necessary conditions as tasks become more complex.
To tackle these issues, the paper proposes a future-predictive success-or-failure classification method called Future-predictive Identifier for Robot Planning (FIRP). FIRP consists of two steps: 1) executing a long-horizon transition prediction of image features obtained by implementing an action plan, and 2) outputting success-or-failure scores with the predicted image features. This two-step method enables success-or-failure classification without executing the action plan, eliminating the need for iterative trials and manual redesign.
The paper also proposes a regularization term called Transition Consistency Regularization (TCR) to improve the long-horizon prediction and classification accuracy. TCR maintains consistency in temporal transition and consistency between actions and transition directions, making the feature distribution more predictable.
The effectiveness of the proposed method is demonstrated through classification and robotic-manipulation experiments. The results show that FIRP achieves competitive classification performance compared to baselines and significantly improves the success rate of long-horizon tasks when combined with a TAMP method.
他の言語に翻訳
原文コンテンツから
arxiv.org
深掘り質問