For decades, the standard paradigm in artificial intelligence and robotics followed a clean, sequential pipeline: perceive, plan, act. An agent would scan its environment, construct an internal model, compute an optimal sequence of actions to reach its goal, and then execute that sequence. This is an open-loop approach. It assumes that the world is a static stage where nothing moves unless the robot moves it, and every action succeeds exactly as intended.
In reality, the physical world is messy, non-deterministic, and highly dynamic. A human might move an object mid-task, a calibration error might cause a gripper to miss by a few millimeters, or an unexpected obstacle might block a path. When an open-loop system encounters these deviations, its rigid plan shatters. Because it lacks real-time feedback during execution, it cannot adapt, leading to catastrophic task failure or frozen, brittle behaviors.
To build truly autonomous systems, the research community has shifted toward closed-loop, reactive execution. Instead of treating planning and execution as separate, chronological phases, modern frameworks tightly couple perception and action, allowing agents to continuously re-evaluate and modify their strategies on the fly.
The Fragility of Open-Loop Execution
To understand why closed-loop execution is necessary, we must first look at where open-loop systems break down. The core vulnerability lies in the decoupling of the initial observation from the final execution phase.
Consider a robotic arm tasked with picking up an object from a cluttered table. In a traditional setup, the robot takes a single snapshot of the scene, plans a trajectory, and replays that trajectory. As Dieter Fox demonstrates in 6-DOF Grasping for Target-driven Object Manipulation in Clutter, relying on a single initial observation makes a system incredibly sensitive to minor calibration errors or physical shifts. If an object is nudged even slightly after the perception phase, the robot will blindly attempt to grasp empty space or collide with surrounding clutter.
This vulnerability scales up to complex manipulation tasks. Edward Johns et al. explore this in On the Effectiveness of Retrieval, Alignment, and Replay in Manipulation, noting that replaying trajectories open-loop prevents recovery if the environment changes post-alignment. Similarly, Yilong Wang notes in One-Shot Dual-Arm Imitation Learning that systems transitioning to open-loop replay after an initial alignment phase cannot adapt if the target moves during the final interaction stage.
Open-Loop: [ Perceive ] ──> [ Plan ] ──> [ Act (Blind Replay) ] ❌ (Fails if world changes)
Closed-Loop: [ Perceive ] ──> [ Plan ] ──> [ Act ]
▲ │
└─────── [ Feedback Loop ] ─┘ ──> (Adapts in real-time)
Whether at the level of low-level motor control or high-level task coordination, open-loop strategies fail because they treat the environment as a fixed equation. As Yixiang Jin et al. point out in RobotGPT: Robot Manipulation Learning from ChatGPT, static initial scans leave an agent entirely unable to recover if objects are displaced before or during a manipulation sequence.
Closing the Loop at the Control Level
The solution to this fragility is to close the loop, transforming execution into a continuous process of online inference and adjustment. This shift from "thinking then acting" to "acting while thinking" dates back to foundational AI robotics work. R. James Firby pioneered this concept in An Investigation into Reactive Planning in Complex Domains by introducing Reactive Action Packages (RAPs), which enabled robots to make decisions based on what is happening now rather than what was predicted beforehand.
Modern architectures implement this reactivity by feeding real-time sensory data directly into the control loop. For vision-based manipulation, Dmitry Kalashnikov et al. developed QT-Opt in Scalable Deep Reinforcement Learning for Vision-Based Robotic Manipulation. By training a reinforcement learning policy on hundreds of thousands of real-world attempts, the system learns to use continuous visual feedback to compensate for physical disturbances, exhibiting emergent behaviors like re-grasping when an object slips.
Instead of computing a single trajectory, reactive systems treat control as an ongoing estimation problem. Dongwon Son formalizes this in Grasping as Inference: Reactive Grasping in Heavily Cluttered Environment, using a particle filter to continuously fuse multi-view observations with closed-loop feedback. This allows the robot to dynamically track and adapt to moving objects in real time.
Recent breakthroughs in generative modeling have further accelerated this capability. The authors of Act to See, See to Act: Diffusion-Driven Perception-Action Interplay for Adaptive Policies introduce a closed-loop "perception-action interplay" using an Action-Guided Diffusion Policy (DP-AG). By utilizing the mathematical gradients of the action-denoising process to refine latent observations in real-time, the policy continuously corrects its path based on immediate environmental feedback.
To balance long-horizon intent with immediate reactivity, researchers often chain coarse global plans with high-frequency local adjustments:
- Granular Re-planning: Zhou Xian et al. introduce ChainedDiffuser: Unifying Trajectory Diffusion and Keypose Prediction for Robotic Manipulation, which combines high-level keypose predictions with fluid, diffusion-based trajectory generation. Closing the loop at a finer granularity allows the robot to handle unexpected object movements during contact-rich tasks.
- Mid-Motion Corrections: Onur Celik et al. argue in Acquiring Diverse Skills using Curriculum Reinforcement Learning with Mixture of Experts that episode-based planning fails during physical collisions because the entire trajectory is determined at the start. Continuous closed-loop policies are required to execute mid-motion corrections.
- Macro-Actions: Panpan Cai et al. propose MAGIC in MAGIC: Learning Macro-Actions for Online POMDP Planning, generating temporally extended action sequences that maintain online reactivity to sudden environmental shifts.
High-Level Semantic Re-planning
Reactivity cannot stop at low-level motor control. If a robot encounters a locked door or drops a vital tool, a smooth motor controller is useless without a high-level strategic shift. Therefore, the closed-loop paradigm must also be integrated into symbolic and semantic reasoning engines.
With the advent of Large Language Models (LLMs) as reasoning engines, researchers have begun wrapping these models in closed-loop architectures to prevent them from hallucinating or executing obsolete instructions. As Svyatoslav Pchelintsev et al. observe in LERa: Replanning with Visual Feedback in Instruction Following, standard LLM planners are often "blind" to the physical world; when an action fails, they stall or continue executing steps that are no longer viable. By feeding visual feedback back into the LLM, the system can dynamically re-plan its high-level logic.
[ High-Level LLM Planner ] ──> Generates Sub-goal ──> [ Low-Level Controller ]
▲ │
│ ▼
[ Environment Feedback ] <─── Updates State ◄─── [ Physical Execution Failure ]
Several frameworks have successfully operationalized this high-level feedback loop:
- Role Decomposition: Haochen Shi et al. introduce OPEx: A Large Language Model-Powered Framework for Embodied Instruction Following, which splits the LLM into an Observer, a Planner, and an Executor. The Observer continuously monitors the environment for errors, allowing the Planner to rewrite the sequence when initial assumptions prove incorrect.
- Self-Correction in Navigation: Xiwen Liang et al. present CorNav in CorNav: Autonomous Agent with Self-Corrected Planning for Zero-Shot Vision-and-Language Navigation, incorporating real-time environmental feedback and trajectory history to allow navigation agents to self-correct and recover from blocked paths.
- Spatio-Temporal Constraints: Wenlong Huang et al. propose ReKep: Spatio-Temporal Reasoning of Relational Keypoint Constraints for Robotic Manipulation. By representing tasks as a sequence of 3D keypoint constraints rather than fixed stages, the robot can dynamically re-plan its high-level strategy—such as choosing to release and regrasp an object—when the initial approach becomes unfeasible.
Scaling to Multi-Robot Teams and Symbolic Spaces
When coordinating multi-robot teams or operating across complex symbolic domains, the coupling between planning and environmental feedback becomes even more pronounced. Debadeepta Dey et al. highlight this fundamental coupling in Efficient Optimization of Control Libraries, noting a tight feedback loop where a robot's action choices directly dictate the obstacles it will perceive in future problem instances.
In multi-agent settings, generating static, upfront coordination plans leaves the entire team vulnerable to single-point failures. Shyam Sundar Kannan et al. address this in SMART-LLM: Smart Multi-Agent Robot Task Planning using Large Language Models, showing that real-time feedback and dynamic re-planning are essential for multi-robot teams to bypass unexpected obstacles like locked doors.
Similarly, Kazuma Obata et al. demonstrate in LiP-LLM: Integrating Linear Programming and dependency graph with Large Language Models for multi-robot task planning that static task dependency graphs cannot recover from dropped objects or human interference without real-time graph updates. To mitigate this, Piyush Gupta et al. implement closed-loop feedback within hierarchical trees in Generalized Mission Planning for Heterogeneous Multi-Robot Teams via LLM-constructed Hierarchical Trees, enabling real-time re-planning and robust error recovery across heterogeneous teams.
Finally, when the environment presents a scenario entirely outside the robot's pre-defined model, the agent must be able to adapt its very understanding of the world. Heng Ji et al. emphasize this in A Language-First Approach for Procedure Planning, demonstrating that mapping visual observations to language allows systems to re-plan on-the-fly when faced with execution failures. When these failures stem from gaps in the robot's underlying capabilities, Julian Förster et al. show in Automatic extension of a symbolic mobile manipulation skill set that autonomous systems can break out of rigid behaviors by exploring new action sequences online, automatically expanding their symbolic knowledge base to adapt to previously unseen tasks.
Why It Works: The Intuition
Closed-loop reactive planning works because it fundamentally re-frames the relationship between the agent and time.
An open-loop planner treats time as a linear track, optimizing for a future that it assumes is perfectly predictable. A closed-loop planner treats time as a series of instantaneous horizons. By constantly sampling the state of the world and treating its current plan as a flexible hypothesis rather than an absolute directive, the agent minimizes the compounding error that naturally occurs when interacting with physical reality.
Whether through high-frequency diffusion gradients at the gripper level or real-time LLM self-correction at the task level, closing the loop ensures that the robot's next action is always grounded in the current state of the world.
Key Foundational Work
An Investigation into Reactive Planning in Complex Domains (1987) — contributed 1 concept elements to this direction.
Integrating Planning and Reacting in a Heterogeneous Asynchronous Architecture for Controlling Real-World Mobile Robots (1992) — contributed 1 concept elements to this direction.