High-Frequency Replanning for Dynamic Robotic Environments

This direction is split into other directions: real-time planning latency. Its last version stays readable and citable here.

Version 1 of 1270 papers276 conceptsfrozen 2026-10-04

In the traditional robotics pipeline, planning is a heavy, deliberate process. A robot senses its environment, builds a map, and then spends seconds—or even minutes—calculating a mathematically optimal path to its goal. This "think-then-act" paradigm works in static factories, but it fails in the messy, moving world of humans. If a pedestrian steps into a robot’s path, a plan that took two seconds to compute becomes obsolete in twenty milliseconds.

The research direction emerging across these papers is a shift toward high-frequency replanning: the ability to update complex trajectories at the millisecond scale. By treating planning not as a one-time event but as a continuous, reactive loop, researchers are enabling robots to navigate crowds, fly aggressively through gaps, and handle shifting objects with fluid, human-like responsiveness.

Breaking the Serial Bottleneck with Parallelism

The primary obstacle to real-time planning is the "serial bottleneck." Traditional sampling-based planners, like RRT*, grow a search tree one branch at a time. This is inherently slow.

To overcome this, Nicolas Perrault and colleagues introduced Kino-PAX: Highly Parallel Kinodynamic Sampling-based Planner, which moves the entire planning process onto the GPU. Instead of growing one branch, it grows thousands of potential paths simultaneously. This massive parallelism allows the system to solve high-dimensional problems in milliseconds rather than seconds. Similarly, researchers have explored using Pixel Processor Arrays (PPAs) to bypass the latency of moving image data to a CPU, as seen in Agile Reactive Navigation for A Non-Holonomic Mobile Robot Using A Pixel Processor Array, enabling robots to react to obstacles at speeds that would overwhelm standard vision systems.

From Scratch to Residuals: Tweaking Instead of Reinventing

If a robot’s environment changes only slightly, it shouldn't need to reinvent its entire path. A key intuition in this field is that local adjustments are cheaper than global searches.

Xi Huang and the team behind MoRe-ERL: Learning Motion Residuals using Episodic Reinforcement Learning demonstrate this through "motion residuals." Instead of generating a new path from scratch, the system learns to "tweak" an existing, stable path to account for new obstacles. This approach combines the reliability of traditional planners with the speed of reinforcement learning.

This philosophy of local adjustment is also present in CSubBT by Huihui Guo and others. By integrating "Conditional Samplers" into behavior trees, the robot can resample specific action parameters—like a grip angle or a foot placement—on the fly when it encounters a failure, avoiding the high computational cost of a full high-level replan.

Generative Composition for Complex Instructions

As robots move into human-centric spaces, they must follow complex, multi-part instructions like "stay on the right while yielding to the pedestrian." Solving these constraints simultaneously in real-time is a massive optimization challenge.

Chen Tang and Zichao Hu address this in ComposableNav by leveraging the mathematical properties of diffusion models. Because diffusion models represent behaviors as gradients, multiple behaviors can be "composed" or blended at runtime. This allows the robot to satisfy novel combinations of commands without needing specific training for every possible scenario, all while maintaining the low latency required for dynamic environments.

Tighter Relaxations and Simplified Physics

Sometimes, the math itself is the bottleneck. Traditional Task and Motion Planning (TAMP) often relies on mixed-integer programs that are notoriously slow to solve.

Xuan Lin and Ye Zhao proposed Logic Network Flow (LNF) to solve this. By encoding temporal logic as network flow constraints rather than tree structures, they produce "tighter" convex relaxations. This mathematical shortcut achieves speedups of several orders of magnitude, making millisecond-level replanning possible even for complex bipedal locomotion.

Similarly, for large-scale swarms, Chang Liu and Silvia Ferrari developed Adaptive Online Distributed Optimal Control (ADOC). Instead of planning for each of the hundreds of robots individually, they treat the swarm as a fluid-like probability distribution. This abstraction turns a nightmare of non-linear programming into a manageable graph-routing problem that can adapt in real-time as new obstacles appear.

Why It Works: The Shift to Reactive Optimization

The common thread across these works is the rejection of the "global optimum" in favor of "sufficiently good" reactive updates. Whether through GPU acceleration, residual learning, or tighter mathematical relaxations, the goal is the same: to reduce the time between perception and action.

When the planning cycle matches the physical velocity of the robot—as argued by Zirui Xu and Vasileios Tzoumas in their work on Resource-Aware distributed Greedy (RAG)—the robot ceases to be a stuttering machine and becomes a truly agile agent capable of navigating the unpredictable world.

Papers in this version

Showing 60 of 270 papers.

History & citationVersion 1 · 2026-10-04

Split into other directions · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_high_frequency_replanning_for_dynamic_ro_5264fc03,
  title = {High-Frequency Replanning for Dynamic Robotic Environments},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/high-frequency-replanning-for-dynamic-robotic-environments/rdr_5264fc03fac044209c8d301f38f5e8a4},
  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_high_frequency_replanning_for_dynamic_ro_5264fc03_v1,
  title = {High-Frequency Replanning for Dynamic Robotic Environments},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/high-frequency-replanning-for-dynamic-robotic-environments/rdr_5264fc03fac044209c8d301f38f5e8a4/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-21},
}