Context-aware LLM-based Safe Control Against Latent Risks

TL;DR

The paper tackles latent risks in autonomous control by introducing a context-aware, multi-layer framework that fuses LLM-based task decomposition, in-context reasoning, and optimization-based control across three distinct timescales.Key contributions include a Subtask generation with Chain-of-Thought module, an MPC-based subtask controller, and a multi-turn, zeroth-order optimization loop that iteratively refines subtask parameters using qualitative and trajectory data.The authors validate the framework through robot-arm pushing and grasping tasks and autonomous-vehicle navigation scenarios, demonstrating improved safety against latent risks without sacrificing efficiency.This work provides a practical blueprint for integrating natural language reasoning with numerical optimization to enable context-sensitive, risk-aware autonomous behavior in dynamic environments.

Abstract

Autonomous control systems face significant challenges in performing complex tasks in the presence of latent risks. To address this, we propose an integrated framework that combines Large Language Models (LLMs), numerical optimization, and optimization-based control to facilitate efficient subtask learning while ensuring safety against latent risks. The framework decomposes complex tasks into a sequence of context-aware subtasks that account for latent risks. These subtasks and their parameters are then refined through a multi-time-scale process: high-layer multi-turn in-context learning, mid-layer LLM Chain-of-Thought reasoning and numerical optimization, and low-layer model predictive control. The framework iteratively improves decisions by leveraging qualitative feedback and optimized trajectory data from lower-layer optimization processes and a physics simulator. We validate the proposed framework through simulated case studies involving robot arm and autonomous vehicle scenarios. The experiments demonstrate that the proposed framework can mediate actions based on the context and latent risks and learn complex behaviors efficiently.

Figures (24)

• Figure 1: Overview of the proposed framework for generating context-aware decisions that anticipate latent risks. The orange, green, and blue shaded areas correspond to the three central design components introduced in Section \ref{['sec:intro']}.
• Figure 2: Workflow of the proposed framework (Algorithm \ref{['alg:proposed_framework']}). The orange part denotes the subtask generation and contextual reasoning in Algorithm \ref{['alg:context-resoning']}. The green part is for optimization-based control in Algorithm \ref{['alg:mpc']}. The blue part illustrates self-adaptive reasoning with zeroth-order optimization in Algorithm \ref{['alg:proposed_framework']}, lines 6-17. The module $\mathtt{LLM}_{\mathtt{coder}}$ utilizes in-context learning, while other LLM components operate in a zero-shot/few-shot setting.
• Figure 3: Three types of feedback loops at different temporal scale.
• Figure 4: Legend illustrating the six approaches compared in our experiments.
• Figure 5: Simulation configurations and results for the robot arm tasks described in Section \ref{['sec:fastadaptation']}. Corresponding design space and legend are shown in Fig. \ref{['fig:legend']}, with the curve colors above matching the respective approaches. In all loss and energy consumption plots, solid dots denote successful task executions, while hollow circles indicate failed attempts.