Robust Control through Receding Horizon Replanning

You are reading version 1 of this direction, frozen 2026-10-04. The latest version is version 2.

Version 1 of 2559 papers634 conceptsfrozen 2026-10-04

In classical robotics and optimization, there is a traditional divide between planning and control. Planning is usually a "one-shot" process: you look at the goal, calculate a long trajectory to get there, and then hand that plan to a controller to execute. However, in the real world—where sensors are noisy, wind gusts push drones off course, and human collaborators move unpredictably—a static plan is a recipe for failure.

The research direction emerging across these papers is a move toward Receding Horizon Replanning. Instead of treating a plan as a fixed script, these researchers treat planning as a continuous, high-frequency process. By planning for the future but only executing the immediate next step before re-evaluating, systems can achieve the robustness of feedback control with the foresight of complex optimization.

The Intuition: Closing the Loop with Open-Loop Tools

The core insight of this field is that you don't necessarily need a complex, global "policy" (a mathematical rule for every possible situation) to handle uncertainty. Instead, you can use "open-loop" optimization—calculating a specific path from point A to point B—and turn it into a "closed-loop" system by doing it over and over again.

Ari Weinstein and Michael Littman demonstrated this with HOLOP (Bandit-Based Planning and Learning in Continuous-Action Markov Decision Processes). They showed that by searching for a sequence of actions but only performing the first one, the system naturally corrects for environmental noise. If the first action doesn't land the agent exactly where it expected, the next planning cycle simply starts from the new, actual position and finds a new path.

This "receding horizon" acts as a buffer against reality. As Jeffrey Mao and Giuseppe Loianno found in Robust Active Visual Perching with Quadrotors on Inclined Surfaces, high-speed drones can land on walls more reliably if they regenerate their trajectories mid-flight. The closer they get to the target, the better their visual data becomes; by replanning, they "wash away" the localization errors that accumulated earlier in the flight.

Efficiency through Local Optimization

A major hurdle for real-time replanning is computational cost. If a robot needs to react in milliseconds, it cannot afford to solve a massive, non-linear optimization problem from scratch every time.

Researchers like Ran Wang argue in On the Search for Feedback in Reinforcement Learning that we should favor "local" approaches. Instead of searching for a global neural network policy, we can periodically re-optimize a local linear law. This is echoed by Sergey Levine and Pieter Abbeel, who use iterative Linear Quadratic Regulators (iLQR) to allow robots to learn manipulation tasks in seconds (Model-based Reinforcement Learning with Parametrized Physical Models and Optimism-Driven Exploration).

To further slash costs, Eivind Bøhn and Tor Arne Johansen proposed using Reinforcement Learning to decide when to replan (Optimization of the Model Predictive Control Update Interval Using Reinforcement Learning). By using a "cheap" controller for minor deviations and only triggering the "expensive" re-optimization when errors exceed a threshold, they reduced computational load by 75% without sacrificing performance.

Handling Multi-Agent and System-Wide Dynamics

The receding horizon framework is particularly powerful when the environment isn't just noisy, but adversarial or collaborative.

In multi-agent settings, Negar Mehr has shown that complex "games" between robots can be simplified into unified optimization problems (Potential iLQR: A Potential-Minimizing Controller for Planning Multi-Agent Interactive Trajectories). By solving these games over a receding horizon, agents can adapt to the unpredictable movements of others in real-time.

This logic scales from individual robots to massive infrastructure:

The Future: Generative and Context-Aware Replanning

The latest frontier in this area involves integrating modern generative AI into the replanning loop. In Generative Trajectory Stitching through Diffusion Composition, Yunhao Luo and Danfei Xu treat planning as a "stitching" task. If an agent drifts off course, a diffusion model "re-stitches" a new trajectory from the current state to the goal.

Similarly, the CAPE framework by Rui Heng Yang uses the unexecuted portion of a previous plan as a "noisy seed" for the next one (CAPE: Context-Aware Diffusion Policy Via Proximal Mode Expansion for Collision Avoidance). This ensures that even as the robot replans at high frequencies to avoid new obstacles, its movements remain smooth and temporally consistent.

By shifting the focus from finding the "perfect" plan to building a system that can "re-plan" perfectly, these researchers are creating autonomous agents that are finally robust enough for the chaos of the real world.

Papers in this version

Showing 60 of 559 papers.

History & citationVersion 2 · 2026-10-04

Current · Permanent link

  1. Version 2 (latest) · 2026-10-04 · 444 papers
  2. Version 1 · 2026-10-04 · 559 papers
Cite this direction

Latest version (a living page):

@misc{lacuna_receding_horizon_control_3ca81ee5,
  title = {receding horizon control},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/receding-horizon-control/rdr_3ca81ee5a28c4d53a3c0b52cfa65e2d8},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_robust_control_through_receding_horizon__3ca81ee5_v1,
  title = {Robust Control through Receding Horizon Replanning},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/robust-control-through-receding-horizon-replanning/rdr_3ca81ee5a28c4d53a3c0b52cfa65e2d8/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}