Future Predictive Success-or-Failure Classification for Long-Horizon Robotic Tasks

TL;DR

The paper addresses the challenge of designing reliable planners for long-horizon robotic tasks by predicting the feasibility of action plans before execution. It introduces FIRP, a future-predictive framework based on a Recurrent State Space Model (RSSM) that forecasts long-horizon image features from an initial scene and an action sequence to produce a success/failure score without executing the plan. To stabilize learning and improve predictive quality, it adds Transition Consistency Regularization (TCR), consisting of Temporal Transition Consistency (TTC) and Action-Transition Consistency (ATC), which shape the feature transitions across time and action categories. Empirical results show FIRP achieves competitive classification performance and, when integrated with a planning method, significantly boosts long-horizon task success rates, demonstrating a practical path to automating condition design in robotics.

Abstract

Automating long-horizon tasks with a robotic arm has been a central research topic in robotics. Optimization-based action planning is an efficient approach for creating an action plan to complete a given task. Construction of a reliable planning method requires a design process of conditions, e.g., to avoid collision between objects. The design process, however, has two critical issues: 1) iterative trials--the design process is time-consuming due to the trial-and-error process of modifying conditions, and 2) manual redesign--it is difficult to cover all the necessary conditions manually. To tackle these issues, this paper proposes a future-predictive success-or-failure-classification method to obtain conditions automatically. The key idea behind the proposed method is an end-to-end approach for determining whether the action plan can complete a given task instead of manually redesigning the conditions. The proposed method uses a long-horizon future-prediction method to enable success-or-failure classification without the execution of an action plan. This paper also proposes a regularization term called transition consistency regularization to provide easy-to-predict feature distribution. The regularization term improves future prediction and classification performance. The effectiveness of our method is demonstrated through classification and robotic-manipulation experiments.

Figures (13)

• Figure 1: Optimization-based planning methods.
• Figure 2: Integration of our method and a planning method.
• Figure 4: Processing pipeline of the proposed method called Future-predictive identifier for robot planning; FIRP. FIRP receives an initial image $I_1$ and actions $\{\mathop{\mathrm{\bf a}}\nolimits_t\}$, then outputs a success score $p$ using long-horizon prediction of subsequent image features $\{\mathop{\mathrm{\bf e}}\nolimits_t\}_{t=2}^T$. The image feature prediction is based on latent variables $\{\mathop{\mathrm{\bf z}}\nolimits_t\}$ behind image features. FIRP is trained using image sequences and three loss functions ${\cal L}_{KL}$, ${\cal L}_{re}$ and ${\cal L}_{ce}$: the first two are for the prediction, and the last is for the classification.
• Figure 5: Pipeline of the future prediction of latent variables $\{\mathop{\mathrm{\bf z}}\nolimits_t\}$. A latent variable $\mathop{\mathrm{\bf z}}\nolimits_t$ consists of a stochastic and deterministic feature and is used to decode an image feature.
• Figure 6: TTC maintains consistency in temporal transition by suppressing the explosive temporal change of image features.