TRATSS: Transformer-Based Task Scheduling System for Autonomous Vehicles

TL;DR

TRATSS addresses single-agent scheduling on graphs by integrating a graph neural network encoder with an autoregressive transformer decoder and reinforcement learning. The three-network decoder selects the next area, the starting point, and the movement pattern, with training guided by REINFORCE to minimize total execution cost measured as distance. The approach is validated in a UAV-style search-and-rescue scenario, demonstrating strong generalization and superior scaling performance compared with the Concorde solver, including robustness across 30–50 area scenarios. This work offers a fast, scalable, and adaptable framework for time-extended scheduling in autonomous systems with complex task dependencies.

Abstract

Efficient scheduling remains a critical challenge in various domains, requiring solutions to complex NP-hard optimization problems to achieve optimal resource allocation and maximize productivity. In this paper, we introduce a framework called Transformer-Based Task Scheduling System (TRATSS), designed to address the intricacies of single agent scheduling in graph-based environments. By integrating the latest advancements in reinforcement learning and transformer architecture, TRATSS provides a novel system that outputs optimized task scheduling decisions while dynamically adapting to evolving task requirements and resource availability. Leveraging the self-attention mechanism in transformers, TRATSS effectively captures complex task dependencies, thereby providing solutions with enhanced resource utilization and task completion efficiency. Experimental evaluations on benchmark datasets demonstrate TRATSS's effectiveness in providing high-quality solutions to scheduling problems that involve multiple action profiles.

Figures (7)

• Figure 1: System workflow.
• Figure 2: TRATSS. (a) The map is normalized and the required features are extracted, (b) The encoder then processes the input features through GNN to obtain a new feature representation, and (c) The decoder processes the new features in 3 networks: the first network selects the next area to visit, the second network assigns the starting point in the chosen area, and finally the third network decides the movement pattern to be followed.
• Figure 3: Available patterns. (a) Vertical Zig-Zag, (b) Horizontal Zig-Zag, and (c) Spiral.
• Figure 4: Pure TSP results.
• Figure 5: Optimality gap for the network choosing the pattern.