Decoupling Planning and Control for Instructable AI Agents

Zineng Tang, Kelsey R. Allen, Sjoerd van Steenkiste, Ishita Dasgupta, Alane Suhr· August 28, 2026 View original

Key takeaways

  • Instruct-to-Act decouples VLM planning from world-model control for AI agents.
  • VLMs provide high-level plans, while controllers handle low-latency actions.
  • Synthetic instruction relabeling trains controllers for language instructability.
  • This approach improves performance and flexibility over integrated VLM-action systems.

Who benefits

RoboticsAutonomous VehiclesLogisticsGamingVirtual Assistants

Summary

Instruct-to-Act is a new system that combines the strengths of VLM planners for high-level instructions with world-model controllers for low-latency actions, decoupling planning and control. By training controllers with synthetic instructions, it outperforms combined VLM-action variants across multiple embodied environments, maintaining fast control and allowing flexible VLM planner integration.

Recent advancements show that instruction-tuned vision-language models (VLMs) excel at generating high-level plans from instructions and observations, but struggle with executing these plans as reliable, low-latency actions in new environments. Conversely, world-model controllers are adept at fast observation-to-action control but lack the ability to follow open-ended task guidance. The new Instruct-to-Act system merges these complementary strengths by decoupling planning from control. It trains a world-model controller to execute autonomously at high frequency, conditioned on sparse, high-level text instructions provided by a VLM planner. To make controllers language-instructable, segments of policy rollouts are relabeled with synthetic instructions, optimizing a behavior-cloning objective alongside reward-maximizing and world-modeling objectives. This decoupled approach consistently outperforms direct VLM action generation and controller-only variants across seven embodied environments, including multi-agent scenarios, while preserving fast control and allowing flexible VLM planner integration without fine-tuning.

Why it matters

This approach offers a more robust and flexible way to build AI agents that can understand complex instructions and execute them reliably in dynamic environments, crucial for robotics and autonomous systems.

How to implement this in your domain

  1. 1Evaluate current agent architectures for tasks requiring both high-level planning and low-latency control.
  2. 2Consider adopting a decoupled VLM-planner and world-model controller approach for new agent development.
  3. 3Explore methods for generating synthetic instructions to train controllers for language instructability.
  4. 4Test the system's performance in environments requiring fast, reactive control alongside complex task execution.
  5. 5Investigate the potential for swapping different VLM planners without retraining the underlying controller.

Original post by Zineng Tang, Kelsey R. Allen, Sjoerd van Steenkiste, Ishita Dasgupta, Alane Suhr

"arXiv:2608.26788v1 Announce Type: new Abstract: Recent work shows that pre-trained, instruction-tuned vision-language models (VLMs) perform well at mapping from instructions and observations to high-level plans, but struggle to realize such plans as reliable low-latency action se…"

View on X

Originally posted by Zineng Tang, Kelsey R. Allen, Sjoerd van Steenkiste, Ishita Dasgupta, Alane Suhr on X · view source

Want to go deeper?

Turn these trends into skills with Learnijoy's hands-on AI & tech courses.

Explore courses

More in AI Engineering & DevTools