FASTFUNNELS: Agile Multi Robot Navigation in Narrow Spaces via Adaptive Dynamic Funnels

Vignesh Rajagopal, Rohan ChandraarXiv 2026 · arXiv

TL;DR

FASTFUNNELS introduces a hierarchical control framework that allows teams of car-like robots to navigate tight, cluttered spaces at high speeds by mimicking the fluid motion of a water droplet. By combining a high-level reinforcement learning "blob" that morphs its shape in real-time with low-level distributed model predictive control, the system achieves nearly 8x the speed of state-of-the-art baselines while maintaining high completion rates.

What Problem Does This Paper Address?

Navigating a single autonomous car at high speed is hard; navigating a team of them through a narrow, obstacle-strewn corridor is a nightmare. Traditionally, researchers use "formation control," where robots maintain a specific geometric shape (like a square or a line).

The problem is that these shapes are usually rigid. When the team hits a bottleneck, they often have to stop or slow down drastically to "switch" from one predefined shape to another (e.g., from a wedge to a single file). This destroys momentum—a critical component for agility. Furthermore, most existing methods struggle with Ackermann steering (the way actual cars turn), which adds complex non-holonomic constraints that simple "unicycle" models ignore. The gap the authors tackle is the "agility-safety" trade-off: how do you keep robots moving fast without them crashing into walls or each other when the space gets tight?

Methods

The authors propose a hierarchical approach called FASTFUNNELS, which separates the "fluidity" of the team from the "safety" of the individual agents.

  1. The Leader (Reinforcement Learning): Instead of a rigid shape, the "leader" is a virtual, deformable elliptical patch (the "funnel"). This patch is trained offline using Proximal Policy Optimization (PPO). It learns to "morph"—stretching, squeezing, and orienting itself—to flow through obstacles like a droplet of water. It doesn't care about individual robot positions; it only cares about finding the fastest, safest path for the "blob."
  2. The Followers (Distributed MPC): Individual robots (the followers) run their own Distributed Model Predictive Control (DMPC). The elliptical patch acts as a "soft forcing function" or a spatial guide. The robots' goal is to stay inside the moving ellipse while avoiding collisions with each other.
  3. Coordination: Robots coordinate by exchanging their planned trajectories rather than just their current states. They use a "separating half-plane" constraint to ensure they don't overlap, effectively carving out their own "slots" within the moving funnel.
Figure 1: Results for a N = 6 shown for 2 (out of 14) benchmark scenes. Each scene is averaged over 20 trials. Remaining scenarios and results can be found in the Appendix. The elliptical funnel patch dynamically adapts its shape and orientation at every control step acting as a soft forcing function to guide the agents fluidly through bottlenecks and narrow gaps at high speed, much like how water flows around rocks and bends. We have tested with N = 6 robots but FASTFUNNELS can scale up to 12 robots as well (beyond N = 12 not tested).

Datasets

The system was evaluated on 14 different benchmark scenarios designed to test specific challenges:

  • Cluttered Environments: Obstacles at narrow hallway entrances.
  • Narrow Corridors: Winding paths that shrink from 9.25m to 3.3m wide.
  • High-Speed Corners: Sharp turns requiring aggressive steering.
  • Slalom: Zig-zagging through protruding bars.

Training and simulation were conducted in the F1TENTH racing simulator, which features high-fidelity car dynamics. The authors also validated the approach using real-world 1/7-scale Traxxas UDR vehicles in a 15m indoor corridor.

Key Insights and Ideas

  • Separation of Concerns: The "aha moment" is delegating agility to a high-level virtual object (the patch) and safety to low-level controllers (DMPC). Trying to optimize for interaction, safety, and fluid agility all at once is computationally overwhelming.
  • Continuous vs. Discrete Adaptation: Unlike the state-of-the-art (DEFORM), which switches between a library of fixed shapes, FASTFUNNELS adapts its shape at 100Hz. This continuous morphing allows the team to maintain high speeds through transitions that would force other systems to halt.
  • The "Soft Forcing" Concept: By treating the formation as a soft boundary rather than a rigid set of target points, the robots have the flexibility to adjust their internal positioning dynamically based on local obstacles.

Key Results

FASTFUNNELS significantly outperformed all baselines in both speed and reliability:

  • Speed: It achieved an average group speed of 7.73 m/s, which is 7.69x faster than the fastest baseline (ORCA).
  • Success Rate: It maintained an 89% completion rate across all scenarios. In contrast, many baselines either crashed at high speeds or were forced to move at "walking speeds" (0.07–1.79 m/s) to finish.
  • Agility: Even at its lowest speed during the tightest bottlenecks, FASTFUNNELS remained faster than the peak speeds of most competitors.
Figure 3: FASTFUNNELS (red curve) at its lowest speed is still several orders of magnitude faster than the second-fastest approach.

The real-world experiments confirmed these findings, with the Traxxas UDRs navigating a cluttered corridor at an average of 3.25 m/s, successfully narrowing their formation to 1.1m to pass through gaps.

Real-world deployment on Traxxas UDR vehicles.

Limitations and Future Work

While impressive, the authors note a few areas for improvement:

  • Information Sharing: The current model assumes robots can share their planned trajectories with each other. In high-interference environments, this might not always be possible.
  • Formal Safety: The "soft" nature of the funnel means safety isn't formally guaranteed (e.g., via Control Barrier Functions), though it is empirically very high.
  • Environment: The current work focuses on static obstacles. Future iterations could extend this to dynamic obstacles (like moving pedestrians) or rough off-road terrains.

Summary updated

Abstract

We focus on the problem of high-speed multi-robot navigation in cluttered spaces with tight spatial constraints. A key challenge is balancing safety with agility, as existing methods often trade one for the other (slowing down near a bottleneck to avoid collisions versus failing to swerve in time at high speeds), especially for car-like systems. Our approach belongs to the family of formation control methods, in which Ackermann-steering agents coordinate while navigating tight spaces. However, current formation control techniques rely on rigid, pre-defined shapes and transitions that depend on initial conditions, the environment, and vehicle dynamics, and are impractical to maintain at high speeds. For example, agents passing through a small gap must first form a square, transform into a line using analytical equations, and then return to a square. In contrast, our approach is inspired by a droplet of water that continuously adapts its shape to the space it flows through without slowing down. We propose a hierarchical algorithm in which, at the high level, a reinforcement learning policy produces a dynamic, blob-like elliptical patch that morphs its shape in real time, squeezing, stretching, and orienting itself around obstacles and through tight spaces. At the low level, the robots are controlled via distributed model predictive control (DMPC) to stay inside the patch, which acts as a soft forcing function, while avoiding collisions with each other. We evaluate our approach, FastFunnels, on simulated F1TENTH racing cars across 14 scenarios with N=1 to N=6 robots, averaged over 20 trials each. FastFunnels achieves an 89% completion rate at 7.73 m/s, while all baselines either fail to complete the track or succeed only at very low speeds (0.07-1.79 m/s, about 40% completion rate).
BibTeX
@misc{rajagopal2026fastfunnels,
  author = {Vignesh Rajagopal and Rohan Chandra},
  title = {{FASTFUNNELS: Agile Multi Robot Navigation in Narrow Spaces via Adaptive Dynamic Funnels}},
  year = {2026},
  eprint = {2611.09005},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.09005},
}
Figure