Real-time goal recognition using approximations in Euclidean space

TL;DR

This work addresses real-time online goal recognition in both continuous and discrete domains under partial observability by sidestepping expensive online planning. It combines a Bayesian inference framework with a fast, approximate trajectory model: offline, it generates multiple approximate trajectories per goal using via points and a polynomial (degree five) trajectory, built from RRT*-derived via points and RL-based optimization to respect dynamics; online, it compares observations to these trajectories via a distance-based likelihood and updates goal posteriors with no planner calls. A key innovation is extending the approach to discrete domains by embedding STRIPS states into vectors and using a Top-k planner to supply multiple optimal plans, enabling unified online inference across domains. Empirically, the method achieves six orders-of-magnitude speedups in the online phase for continuous domains and competitive accuracy vs state-of-the-art methods, with performance improving as the number of trajectory solutions k increases (up to a point). The results demonstrate a practical pathway to real-time, cross-domain online goal recognition for robotics applications that require sub-second responsiveness.

Abstract

While recent work on online goal recognition efficiently infers goals under low observability, comparatively less work focuses on online goal recognition that works in both discrete and continuous domains. Online goal recognition approaches often rely on repeated calls to the planner at each new observation, incurring high computational costs. Recognizing goals online in continuous space quickly and reliably is critical for any trajectory planning problem since the real physical world is fast-moving, e.g. robot applications. We develop an efficient method for goal recognition that relies either on a single call to the planner for each possible goal in discrete domains or a simplified motion model that reduces the computational burden in continuous ones. The resulting approach performs the online component of recognition orders of magnitude faster than the current state of the art, making it the first online method effectively usable for robotics applications that require sub-second recognition.

Figures (11)

• Figure 1: Example of trajectories generated in different stages of the experimental scenario.
• Figure 2: Conditional probabilities $P(\hat{m}_{p_1}^{g_n} \mid O_{p_1}^{g_2})$ for all goals in problems starting at $p_1$ and goal point at $g_2$.
• Figure 3: Comparison between on-line goal recognition methods.
• Figure 4: Aftershock, Archipelago, BigGameHunters, and Brushfire scenarios used in the experiments.
• Figure 5: Caldera, CrescentMoon, Desolation, and EbonLakes scenarios used in the experiments.