Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Toward Agentic AI: Task-Oriented Communication for Hierarchical Planning of Long-Horizon Tasks

Sin-Yu Huang

TL;DR

This work tackles bandwidth-efficient execution of long-horizon tasks by enabling hierarchical, task-focused communication in agentic AI systems. It introduces HiTOC, a framework where a high-level planner and a low-level actor on an edge server coordinate with a robot via subtask-conditioned information and joint source-channel coding. A conditional variational information bottleneck (cVIB) objective jointly trains the subtask encoder, JSCC encoder/decoder, and the action predictor to preserve action-relevant information while minimizing task-data and subtask information rates. Experiments on AI2-THOR with MAP-THOR demonstrate that HiTOC outperforms three baselines in task success, achieving better subtask feature preservation and substantial bitrate reductions compared to non-task-conditioned schemes. Overall, the approach offers a principled, scalable path to efficient, edge-based task execution for complex, long-horizon robotic tasks.

Abstract

Agentic artificial intelligence (AI) is an AI paradigm that can perceive the environment, reason over observations, and execute actions to achieve specific goals. Task-oriented communication supports agentic AI by transmitting only the task-related information instead of full raw data in order to reduce the bandwidth requirement. In real-world scenarios, AI agents often need to perform a sequence of actions to complete complex tasks. Completing these long-horizon tasks requires a hierarchical agentic AI architecture, where a high-level planner module decomposes a task into subtasks, and a low-level actor module executes each subtask sequentially. Since each subtask has a distinct goal, the existing task-oriented communication schemes are not designed to handle different goals for different subtasks. To address this challenge, in this paper, we develop a hierarchical task-oriented communication (HiTOC) framework. We consider a system with an edge server and a robot as an edge device. The high-level planner and low-level actor modules reside on the edge server. The robot transmits only the environment information that is relevant to the current subtask in order to complete a long-horizon task. We propose a conditional variational information bottleneck (cVIB) approach to train the HiTOC framework to adaptively transmit minimal information required for each subtask. Simulations conducted on the AI2-THOR platform demonstrate that the proposed HiTOC framework outperforms three state-of-the-art schemes in terms of the success rate on MAP-THOR benchmark.

Toward Agentic AI: Task-Oriented Communication for Hierarchical Planning of Long-Horizon Tasks

TL;DR

This work tackles bandwidth-efficient execution of long-horizon tasks by enabling hierarchical, task-focused communication in agentic AI systems. It introduces HiTOC, a framework where a high-level planner and a low-level actor on an edge server coordinate with a robot via subtask-conditioned information and joint source-channel coding. A conditional variational information bottleneck (cVIB) objective jointly trains the subtask encoder, JSCC encoder/decoder, and the action predictor to preserve action-relevant information while minimizing task-data and subtask information rates. Experiments on AI2-THOR with MAP-THOR demonstrate that HiTOC outperforms three baselines in task success, achieving better subtask feature preservation and substantial bitrate reductions compared to non-task-conditioned schemes. Overall, the approach offers a principled, scalable path to efficient, edge-based task execution for complex, long-horizon robotic tasks.

Abstract

Agentic artificial intelligence (AI) is an AI paradigm that can perceive the environment, reason over observations, and execute actions to achieve specific goals. Task-oriented communication supports agentic AI by transmitting only the task-related information instead of full raw data in order to reduce the bandwidth requirement. In real-world scenarios, AI agents often need to perform a sequence of actions to complete complex tasks. Completing these long-horizon tasks requires a hierarchical agentic AI architecture, where a high-level planner module decomposes a task into subtasks, and a low-level actor module executes each subtask sequentially. Since each subtask has a distinct goal, the existing task-oriented communication schemes are not designed to handle different goals for different subtasks. To address this challenge, in this paper, we develop a hierarchical task-oriented communication (HiTOC) framework. We consider a system with an edge server and a robot as an edge device. The high-level planner and low-level actor modules reside on the edge server. The robot transmits only the environment information that is relevant to the current subtask in order to complete a long-horizon task. We propose a conditional variational information bottleneck (cVIB) approach to train the HiTOC framework to adaptively transmit minimal information required for each subtask. Simulations conducted on the AI2-THOR platform demonstrate that the proposed HiTOC framework outperforms three state-of-the-art schemes in terms of the success rate on MAP-THOR benchmark.
Paper Structure (6 sections, 26 equations, 5 figures, 1 table, 1 algorithm)

This paper contains 6 sections, 26 equations, 5 figures, 1 table, 1 algorithm.

Table of Contents

  1. Introduction
  2. The Proposed Hierarchical Task-Oriented Communication Framework
  3. The conditional variational information bottleneck approach
  4. Model Architecture and Training Procedure
  5. Performance Evaluation
  6. Conclusion

Figures (5)

  • Figure 1: The system model of our proposed HiTOC framework.
  • Figure 2: The probabilistic graphical model of the proposed HiTOC framework.
  • Figure 3: The architecture of the conditional module, JSCC encoder, and JSCC decoder.
  • Figure 4: Comparison of task-specific images from different baselines under Rayleigh fading channel with SNR = 0 dB for the subtasks (a) "Navigate to the tomato" and (b) "Locate the fridge." In (v), JPEG2000 is marked as Not Applicable as it is unable to recover the image.
  • Figure 5: Comparison of the success rate versus SNR under (a) AWGN and (b) Rayleigh fading channel models across different baselines.