EVENT DOSSIER
The researchers proposed a schema-constrained language model aimed at improving robot strategies without disrupting local learning. This approach was tested in a three-robot system implemented in NetLogo-Python, where the robots shared motion dynamics but used different LLM backends. The system adopted an architecture with a decentralized LLM to generate team actions, restricting LLM inference to回合-level strategy generation and refinement rather than tick-level action selection. Each robot communicated across LLMs through shared回合 summaries containing strategies, results, and learning feedback. UCB performed mode selection, and the strategy-conditioning Double DQN carried out tick-level action selection based on navigation variables. In a fixed simulation environment, this complete configuration successfully reached the target in all 90 robot-rotations records, achieving the lowest intermediate completion time (42 ticks) and the 90th percentile (73.2 ticks).
本文提出一种受 Schema 约束的语言模型，用于在不破坏局部学习的情况下完善机器人策略。在 NetLogo-Python 实现中，三台共享运动动力学但使用不同 LLM 后端的机器人独立结合 LLM 策略代理、UCB 带子和 Double DQN 控制器，且无中心 LLM 生成团队动作。LLM 推理被限制在回合级策略生成与完善，而非 tick 级动作选择；机器人通过包含策略、结果和学习反馈的共享回合摘要进行跨 LLM 通信。UCB 执行完善模式选择，而策略条件 Double DQN 基于导航变量、活跃策略参数和 LLM 动作先验执行 tick 级动作选择。在固定仿真中，完整配置在全部 90 个机器人 - 回合记录中均到达目标，实现了最低的中间完成时间（42 ticks）和第 90 百分位数（73.2 tick…