M2ED:面向长时序具身任务的主动求助元认知多模态决策框架OA

Embodied intelligent agents often fail in long-horizon tasks due to perceptual uncertainty combined with complex semantic structure in real environments.Under electromagnetic spectrum constraints,spectrum occupancy and in-terference can cause missing observations and link latency,further weakening the robustness of long-horizon decision-mak-ing.In this context,the agent should not only trigger help-seeking at critical moments,but also dynamically assess when to ask questions based on time-varying risk,and continuously verify residual uncertainty after receiving assistance.Most exist-ing active querying mechanisms rely on static uncertainty criteria and assume immediate resumption of execution after a query,lacking the ability to jointly model environmental semantic risk and internal cognitive state,as well as to perform continual verification during inference.To address these limitations,we propose a metacognitive and multimodal embodied decision framework(M2ED).First,we build a task-oriented semantic environment complexity extraction module on top of a general-purpose vision-language model,which converts the current observation and dynamic instructions into structured measures of entity density,relational complexity,and descriptive entropy,forming an interpretable semantic complexity vector.Second,we design a metacognitive uncertainty-guided querying controller(MUQC),which takes the semantic com-plexity vector and policy perplexity as joint inputs.The controller is learned via a two-stage procedure,consisting of offline pretraining with soft-label supervision and online fine-tuning with proximal policy optimization(PPO),enabling adaptive query timing with fewer queries in simple scenarios and more queries in complex ones.Finally,we introduce an iterative metacognitive verification loop(IMVL)at inference time.Once a query is triggered,the agent pauses task execution and performs multi-round constrained perception and continuous risk calibration,resuming action execution only when both the policy perplexity and the risk score induced by the semantic complexity vector fall into a low-risk regime,thereby suppress-ing the accumulation and propagation of residual uncertainty over long horizons.Experiments on unseen environments in the action learning from realistic environments and directives(ALFRED)embodied benchmark show that M2ED improves task success to 19.6% while maintaining efficient interaction,limiting the average number of questions to 4.9,and demon-strating strong cross-scene generalization and long-horizon robustness.