Closed Loop Planning and Reactive Replanning

652 concepts458 papers6 proposals

Imagine trying to navigate a crowded, bustling room with your eyes closed after looking at a map just once. Any slight disturbance—a person stepping into your path, a slippery patch on the floor, or an unexpected obstacle—will throw you completely off course.

Traditional automated planners often suffer from this exact vulnerability. They operate in an "open-loop" fashion, generating a complete sequence of actions upfront based on an initial snapshot of the world. However, as Lin et al. point out, static, open-loop plans inevitably fail because they cannot account for execution errors or real-time environmental shifts.

To build truly resilient autonomous systems—whether they are physical robots manipulating objects or digital agents navigating software interfaces—we must transition from static execution to continuous adaptation. This paradigm is known as closed-loop planning or reactive replanning. As summarized in a comprehensive survey by Wang et al., closed-loop planning incorporates continuous external or internal feedback to iteratively update an agent's strategy during task execution, enabling robust error correction in complex, multi-step scenarios.

The Anatomy of a Closed-Loop Plan

Instead of treating planning as a one-time pre-computation step, closed-loop systems turn planning into a continuous runtime process. The system operates in a tight, iterative cycle: Perceive, Reason, Act, and Reflect.

       [ Initial Goal ]
              │
              ▼
       ┌──────────────┐
  ┌───>│   Plan/Act   │
  │    └──────┬───────┘
  │           │
  │           ▼
  │    ┌──────────────┐
  │    │   Perceive   │
  │    └──────┬───────┘
  │           │
  │           ▼
  │    ┌──────────────┐
  │    │   Reflect    │──(Discrepancy Detected?)──> [ Re-plan / Adjust ]
  │    └──────────────┘                                      │
  │                                                          │
  └──────────────────────────────────────────────────────────┘

Consider the SayCan framework by Ichter et al.. Rather than executing a long-horizon kitchen task blindly, the robot evaluates both its task progress and the environment's immediate physical affordances after every single skill execution. This allows the system to dynamically calculate the best next step based on the actual state of the world, recovering gracefully if an action slips or fails.

This concept is a foundational pillar of autonomy. Decades ago, The Navigation System of the JPL Robot pioneered a reactive loop that aborted pre-planned paths and triggered immediate replanning upon encountering unexpected obstacles, proving that real-time environmental feedback is essential for navigating unmapped terrain.

Discrepancy Detection and Execution Monitoring

The core engine driving reactive replanning is execution monitoring—the mechanism that determines when a plan needs to change. The agent must constantly compare its expectations against reality.

The authors of ATLAS (Cheng et al.) formalize this as discrepancy-triggered replanning. The agent uses an internal world model to predict the outcome of its actions. If the observed state of the environment diverges significantly from this prediction, the agent halts its current execution sequence and regenerates its plan to match the new reality.

Similarly, Chen et al. implement a continuous observation loop that uses temporal projection to verify plan validity. If a human moves an object mid-task, creating a contradiction between the robot's internal beliefs and its sensory data, the system triggers an immediate replan. This approach serves as a highly efficient alternative to exhaustive contingency planning; instead of wasting computational power pre-calculating every possible future scenario, the agent simply plans for the most likely path and reacts dynamically when a discrepancy arises.

Hierarchies of Adaptation

When a closed-loop system detects a failure or an environmental shift, it does not always need to throw away its entire strategy. Adaptation operates across a hierarchy of severity:

1. Local Parameter Tuning

Minor glitches do not require full cognitive replanning. As Karapinar et al. demonstrate, when a low-level action fails (such as a robot slipping while trying to grasp an object), the system first attempts to resolve the issue by adjusting continuous parameters—like changing the approach angle or orientation—treating the failure as a temporary execution hiccup rather than a structural flaw in the plan.

2. Subgoal and Path Switching

If an obstacle completely blocks a chosen path, the agent must alter its intermediate milestones. Guo introduces a timeout and switching mechanism for complex temporal logic tasks. If a specific subgoal becomes unreachable due to environmental constraints, the agent dynamically re-evaluates its path through a task automaton and switches to an alternative sequence of subgoals to satisfy the overarching objective.

3. Full State Re-evaluation

When the environment changes drastically, the agent must rebuild its problem representation from scratch. Barrios shows that by treating execution failures as triggers for entirely new deliberation cycles, an agent can generate a fresh problem definition based on the current state and invoke its planner to find a brand-new path to the goal, often resolving complex disruptions in fractions of a second.

Convergence with Modern Foundation Models

While early reactive systems relied on rigid symbolic logic, the modern resurgence of this idea leverages Large Language Models (LLMs) and Vision-Language Models (VLMs) to handle open-world perception and reasoning.

In digital environments, such as smartphone graphical user interface (GUI) automation, Zhang et al. introduce Dynamic Planning of Thoughts (D-PoT). Instead of relying on a static, pre-computed chain of actions, the agent maintains an evolving execution log updated via real-time visual feedback. If a website or app behaves unexpectedly, the agent refines its future steps on the fly, preventing the cascading errors that typically derail static automation scripts.

In physical manipulation, Mo et al. utilize a Planner-Critic architecture (CLEA) that continuously monitors execution, allowing open-source foundation models to detect and prune outdated or redundant actions in real-time.

Furthermore, this closed-loop reactivity can emerge directly within low-level control policies. By combining iterative action refinement with receding horizon control, the Diffusion Policy framework developed by Chi et al. continuously replans action sequences from the current state. This creates an emergent robustness: if an object is violently nudged or moved mid-task, the policy naturally corrects its trajectory without needing explicit error-recovery training data.

Ultimately, whether applied to a planetary rover in 1977 or a multimodal web agent today, the core intuition remains identical: true intelligence lies not in creating a flawless, rigid blueprint, but in the ability to maintain a continuous, reactive conversation with an unpredictable world.

History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_dynamic_plan_adaptation_650ddffd,
  title = {dynamic plan adaptation},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/dynamic-plan-adaptation/rdr_650ddffdf9724d1eb2daad39de1e2b6b},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-21},
}

This exact version (version 1, frozen text):

@misc{lacuna_dynamic_plan_adaptation_650ddffd_v1,
  title = {dynamic plan adaptation},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/dynamic-plan-adaptation/rdr_650ddffdf9724d1eb2daad39de1e2b6b/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-21},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake