TinyMA-IEI-PPO: Exploration Incentive-Driven Multi-Agent DRL with Self-Adaptive Pruning for Vehicular Embodied AI Agent Twins Migration

TL;DR

The paper tackles real-time migration of Vehicle Embodied AI Twins (VEAATs) in Vehicular Embodied AI Networks (VEANETs) under constrained edge resources. It introduces a novel MLMF Stackelberg incentive mechanism to model AVs and RSUs interactions and a lightweight TinyMA-IEI-PPO algorithm that uses self-adaptive dynamic structured pruning guided by individual exploration incentives to run on computation-constrained AVs. Key contributions include a virtual immersive QoE-aware utility model, formal Stackelberg equilibrium analysis with existence and uniqueness proofs, and a self-adaptive pruning framework that preserves near-SE performance while reducing neural complexity. The approach demonstrates convergence near the Stackelberg Equilibrium and effective pruning under sparse regimes, offering scalable deployment for edge-assisted embodied AI in dynamic VEANET environments with practical implications for autonomous driving and edge resource management.

Abstract

Embodied Artificial Intelligence (EAI) addresses autonomous driving challenges in Vehicular Embodied AI Networks (VEANETs) through multi-modal perception, adaptive decision-making, and hardware-software co-scheduling. However, the computational demands of virtual services and the inherent mobility of autonomous vehicles (AVs) necessitate real-time migration of Vehicular Embodied Agent AI Twins (VEAATs) between resource-constrained Roadside Units (RSUs). This paper proposes a novel framework for efficient VEAAT migration in VEANETs, combining a multi-leader multi-follower (MLMF) Stackelberg game-theoretic incentive mechanism with a tiny multi-agent deep reinforcement learning (MADRL) algorithm. First, We propose an virtual immersive experience-driven utility model that captures AV-RSU dynamic interactions by integrating AVs' social influence, service complementarity and substitutability, and RSUs' resource allocation strategies to optimize VEAAT migration decisions. Second, to enhance training efficiency and enable efficient deployment on computation-constrained AVs while preserving exploration-exploitation performance, we propose TinyMA-IEI-PPO, a self-adaptive dynamic structured pruning algorithm that dynamically adjusts neuron importance based on agents' exploration incentives. Numerical results demonstrate that our approach achieves convergence comparable to baseline models and closely approximates the Stackelberg equilibrium.

Figures (10)

• Figure 1: The system model for VEAAT migration.
• Figure 2: TinyIEI-MAPPO algorithm’s Architecture for the VEAAT migration.
• Figure 3: Comparison of episode reward curves of TinyMA-IMI-PPO and baselines for the MFML Stackelberg Game.
• Figure 4: Utilities of AVs and RSUs
• Figure 5: The pricing strategies of RSUs and the bandwidth demands of AVs

Theorems & Definitions (5)

• Lemma 1
• Theorem 1
• Lemma 2
• Theorem 2
• Theorem 3