GRAML: Goal Recognition As Metric Learning

TL;DR

GRAML tackles Online Dynamic Goal Recognition by reframing GR as metric learning. It trains a Siamese LSTM to embed observation traces into a space where traces toward the same goal are close and those toward different goals are distant, enabling one-shot adaptation for new goals via a goal-adaptation phase. The framework offers two strategies, BG-GRAML and GC-GRAML, for domain learning and data generation, supporting both discrete and continuous environments, and uses self-supervised generation to build robust embeddings without needing solved GR tasks. Empirically, GRAML achieves competitive GR accuracy with favorable adaptation and inference times across six diverse domains, highlighting its practical potential for dynamic, real-world GR and ODGR tasks.

Abstract

Goal Recognition (GR) is the problem of recognizing an agent's objectives based on observed actions. Recent data-driven approaches for GR alleviate the need for costly, manually crafted domain models. However, these approaches can only reason about a pre-defined set of goals, and time-consuming training is needed for new emerging goals. To keep this model-learning automated while enabling quick adaptation to new goals, this paper introduces GRAML: Goal Recognition As Metric Learning. GRAML uses a Siamese network to treat GR as a deep metric learning task, employing an RNN that learns a metric over an embedding space, where the embeddings for observation traces leading to different goals are distant, and embeddings of traces leading to the same goals are close. This metric is especially useful when adapting to new goals, even if given just one example observation trace per goal. Evaluated on a versatile set of environments, GRAML shows speed, flexibility, and runtime improvements over the state-of-the-art GR while maintaining accurate recognition.

Figures (6)

• Figure 1: Depiction of ODGR intervals and the inputs at each (left) and a rough approximation of the time spent on each phase by representative GR frameworks (right).
• Figure 2: GRAML's implementation for the different phases of an ODGR problem. The numbers at the top left of each box reflect the section in which each component is discussed.
• Figure 3: The evaluation environments from top-left to bottom-right: Minigrid-SimpleCrossing, Minigrid-LavaCrossing, Parking, PointMaze-Obstacle, PointMaze-FourRooms and Panda-Gym.
• Figure 4: Confusion matrices for plan similarity and recognition confidence in the inference phase for GRAML across multiple tasks in the PointMaze environment where the goals differ from the set of base goals from the domain learning phase.
• Figure 5: BG-GRAML and GC-GRAML accuracy in Parking.

Theorems & Definitions (3)

• Definition 1
• Definition 2
• Definition 3