💻 computer science

MinNav: Minimalist Navigation Using Optical Flow For Active Tiny Aerial Robots

This paper introduces MinNav, a minimalist navigation stack for tiny aerial robots that leverages monocular optical flow and active exploration to autonomously navigate through environments with static and dynamic obstacles and unknown gaps without prior scene knowledge, achieving a 70% success rate with significantly lower computational requirements than depth-based methods.

Original authors: Aniket Patil, Mandeep Singh, Uday Girish Maradana, Nitin J. Sanket

Published 2026-06-09

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Aniket Patil, Mandeep Singh, Uday Girish Maradana, Nitin J. Sanket

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine a tiny drone, the size of a hummingbird, trying to fly through a chaotic forest. It has no map, no GPS, and no 3D scanner to see how far away trees or birds are. All it has is a single camera eye and a very small brain. This is the challenge the paper "MinNav" tackles.

Here is the story of how they solved it, using simple analogies.

The Problem: Flying Blind in a Busy World

Most robots navigate by building a detailed 3D map of the world, like a human looking at a room and mentally measuring the distance to every chair. But for a tiny drone, carrying the heavy computer needed to build that map is impossible. It's like asking a bicycle to carry a truck engine.

Furthermore, the world is messy. There are:

Static obstacles: Trees, rocks, and walls that don't move.
Dynamic obstacles: Falling leaves, birds, or other moving things.
Unknown gaps: Weirdly shaped holes in the middle of a wall that the robot has never seen before.

The researchers wanted a way for the drone to fly through all of this using only a single camera, without knowing what was coming next.

The Solution: The "Bee-Like" Strategy

The team, led by Aniket Patil and Nitin Sanket, created a system called MinNav. Instead of trying to measure exact distances (which is hard and slow), they used Optical Flow.

The Analogy: The Wind in Your Face
Imagine you are running down a hallway. If you look at the walls, the closer the wall is, the faster the paint seems to rush past your eyes. If the wall is far away, the paint moves slowly. This "rush of images" is optical flow.

High flow speed = Something is very close (a wall).
Low flow speed = Something is far away (open space).

However, there's a catch. If the drone spins around, the whole world blurs, making it hard to tell what is close and what is far. Also, if the drone looks straight at a gap, the "rush" stops right in the middle of its vision, creating a "dead spot" where it can't see anything.

The Secret Sauce: "Wandering" to See Better

To fix the dead spots and the spinning confusion, the drone doesn't just fly straight. It uses an Active Strategy.

The Analogy: The Peering Bee
Think of a bee trying to fly through a flower. It doesn't just zoom straight in; it wiggles its head side-to-side (peering) to get a better look.
MinNav does the same thing. It makes tiny, controlled "wiggles" or exploratory movements side-to-side.

Why? These wiggles create a little bit of motion in the camera's view, which fills in the "dead spots" and helps the drone figure out where the open space is.
The Uncertainty Check: The drone also uses a special "uncertainty detector." If a moving object (like a bird) flies by, the math gets "confused" (high uncertainty). The drone uses this confusion to know, "Oh, something is moving fast and I need to dodge it!"

How It Works Step-by-Step

The Wiggle: The drone wiggles slightly to gather information.
The Scan: It looks for the area where the "image rush" is slowest (meaning the most open space).
The Move: It steers toward that open space.
The Goal: It keeps moving forward toward its destination.
The Emergency Brake: If it sees something moving fast toward it (high uncertainty + high speed), it immediately dodges, prioritizing safety over the path.

The Results: Good Enough for the Wild?

The researchers tested this on a real drone in a netted room filled with cardboard boxes, swinging milk jugs (to simulate moving obstacles), and weirdly shaped holes in foam walls.

Success Rate: The drone successfully navigated through these messy scenarios 70% of the time.
Efficiency: While it didn't fly the absolute shortest path (because it was "wiggling" to see), it was incredibly fast at processing information. It used 9 times less computing power than other advanced methods that try to build 3D maps.

The Big Takeaway

This paper proves that you don't need a supercomputer or a 3D laser scanner to fly a tiny robot through a messy, unpredictable world. By mimicking how insects "wiggle" to see, and by using the simple "rush" of images on a camera, a tiny robot can figure out where to go, dodge moving dangers, and fly through unknown gaps.

It's not perfect yet (it failed 30% of the time, mostly because it couldn't see things coming from the side), but it's a major step toward giving tiny, cheap drones the ability to fly autonomously in the real, chaotic world.

Technical Summary: MinNav – Minimalist Navigation for Active Tiny Aerial Robots

Problem Formulation

The paper addresses the challenge of autonomous navigation for tiny aerial robots (specifically quadrotors) in unstructured, "wild" environments. These environments contain a mix of static obstacles, dynamic obstacles (e.g., falling leaves, moving branches), and unknown-shaped gaps.

The core constraints are:

Sensor Limitations: The robot relies solely on a front-facing monocular camera for high-level perception. Down-facing sensors (optical flow, laser altimeter, IMU) are restricted to low-level attitude and altitude control.
Computational Constraints: All perception and decision-making must occur onboard the robot using embedded hardware (e.g., Jetson Orin Nano), precluding heavy 3D reconstruction or cloud processing.
Lack of Prior Knowledge: The system possesses no prior information regarding the scene's geometry, the ordering of obstacles, the size or shape of gaps, or the locations of dynamic objects.
Dynamic Complexity: The system must distinguish between static and dynamic obstacles and react to dynamic threats in real-time without explicit 3D motion modeling.

Existing depth-based methods often fail due to sensor weight, power consumption, or environmental conditions (e.g., direct sunlight, dust) that hinder active sensors. Conversely, existing structureless (optical flow) methods often struggle with the "dead spot" near the Focus of Expansion (FOE), rotational flow interference, and distinguishing dynamic obstacles from static ones without prior scene knowledge.

Methodology

The proposed solution, MinNav, is a unified navigation stack that combines optical flow, flow uncertainty, and active perception (exploratory motion) to navigate without reconstructing the 3D scene.

1. Perception: Optical Flow and Uncertainty

Instead of estimating metric depth, MinNav utilizes a lightweight, multi-scale pyramidal neural network to predict:

Optical Flow ( $\dot{p}$ ): The apparent motion of pixels.
Uncertainty ( $\Upsilon$ ): A measure of confidence in the flow prediction.

Scene Motif Classification:
The system classifies the environment into three primary motifs based on flow and uncertainty cues:

Static Obstacles & Unknown Gaps: Identified by regions of low optical flow magnitude (indicating free space) surrounded by high flow. The system targets the largest contour of low flow to navigate through gaps.
Dynamic Obstacles: Identified by a combination of high optical flow magnitude and high uncertainty. The high uncertainty arises from occlusion and rapid changes in the scene caused by independent object motion. This allows the system to distinguish moving threats from static obstacles that might also have high flow due to proximity.

Network Architecture:

A fully convolutional network with 5 levels (coarse-to-fine) predicting incremental flow.
Trained on FlyingChairs2 and fine-tuned on FlyingThings3D.
Efficiency: 2.8M parameters, capable of running on a Jetson Orin Nano with ~68ms inference time (640x480 resolution).

2. Control: Active Perception Strategy

To overcome the limitations of optical flow (specifically the low flow magnitude near the FOE and rotational flow interference), MinNav employs an active control policy inspired by honey bee "peering" behavior. The control loop consists of three iterative steps:

Diagonal Exploratory Motion: The robot performs small, open-loop diagonal movements in the X-Y plane. This shifts the FOE away from the image center and introduces parallax, generating sufficient flow data to reason about the scene structure without relying on metric depth.
Alignment to Free Space: Based on the flow data gathered during exploration, the robot calculates the center of the largest low-flow region (free space) and executes a velocity command to align with it.
Global Goal Alignment: The robot orients its heading toward the global goal direction (Z-axis) and moves forward.

Dynamic Obstacle Dodging:
A separate high-priority thread monitors for dynamic obstacles (high flow + high uncertainty). Upon detection, an interrupt triggers an immediate acceleration-based dodging maneuver to minimize latency, prioritizing collision avoidance over path optimality.

Key Contributions

Unified Framework: The first solution to handle static obstacles, dynamic obstacles, and unknown-shaped gaps simultaneously using a single structureless formulation without prior scene knowledge.
Lightweight Network: A novel multi-scale optical flow and uncertainty network optimized for embedded tiny aerial robots.
Active Control Policy: A formulation that integrates exploratory motion with goal-directed navigation, effectively mitigating optical flow limitations (FOE dead spots) and enabling robust scene understanding.
Real-World Validation: Successful demonstration on a custom quadrotor (Corgi210) in complex real-world scenarios involving randomized obstacle ordering and dynamic threats.

Experimental Results

The authors evaluated MinNav in both real-world and simulation environments.

Real-World Experiments

Setup: 4 randomized scenarios in a netted facility involving static cardboard obstacles, dynamic swinging objects (milk jugs, boxes), and foamcore gaps.
Performance: Achieved a 70% success rate (56/80 trials).
Failure Analysis: Most failures occurred when dodging dynamic obstacles, attributed to the lack of omnidirectional sensing (hitting side nets) and latency in the detection pipeline (~100ms).

Simulation Experiments

Comparison: MinNav was compared against depth-based methods (MorphEyes, MiDAS) and other optical flow baselines (RAFT).
Success Rate: MinNav with exploration achieved 82% success in simulation, comparable to RAFT (84%) and MiDAS-v3.1 (88%), but significantly outperforming MiDAS-v2.1S (38%) which struggled with dynamic obstacles.
Efficiency: MinNav is 9× faster in inference time (68.2ms vs. 615.3ms for RAFT) and requires significantly fewer FLOPs (27.14 G vs. 211.01 G) and parameters (2.8M vs. 5.2M).
Path Length: While depth-based methods yield the shortest paths, MinNav's path length increase is manageable given the massive reduction in computational cost.

Significance and Claims

The paper claims that MinNav represents a significant step toward minimalist autonomy for tiny aerial robots. Its primary significance lies in:

Feasibility on Tiny Platforms: Demonstrating that complex navigation in unstructured, dynamic environments is possible without heavy 3D sensors or massive compute, relying instead on efficient optical flow and active sensing.
Robustness to Dynamics: Successfully distinguishing and reacting to dynamic obstacles using only monocular flow uncertainty, a capability often missing in prior structureless approaches.
Complementary Role: The authors position MinNav not necessarily as a replacement for depth-based systems, but as a highly efficient, low-power alternative or a robust "backup" system when depth sensors fail or are unavailable.

The work suggests that by leveraging the active nature of the robot (moving to see better), tiny agents can overcome the inherent ambiguities of monocular vision, bringing large-scale deployment of insect-sized robots in the wild closer to reality.