💻 computer science

OmniPlanner: Universal Exploration and Inspection Path Planning across Robot Morphologies

This paper introduces OmniPlanner, a unified and modular path planning framework that enables autonomous aerial, ground, and underwater robots to perform robust exploration and inspection across diverse unstructured environments by integrating volumetric and viewpoint-based strategies with a platform abstraction layer for minimal cross-domain retuning.

Original authors: Angelos Zacharia, Mihir Dharmadhikari, Mohit Singh, Kostas Alexis

Published 2026-03-05

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Angelos Zacharia, Mihir Dharmadhikari, Mohit Singh, Kostas Alexis

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 you are the captain of a fleet of very different vehicles: a drone that flies, a four-legged robot dog that walks, and a submarine that swims. Your goal is to explore a mysterious, dark, and dangerous place (like a collapsed mine, a dense forest, or a sunken ship) and map it out, while also taking close-up photos of specific things you find.

The problem with most current robot software is that it's like having three different captains who only speak three different languages. The drone captain doesn't know how to drive the robot dog, and the submarine captain has no idea how to fly the drone. If you want to switch robots, you have to throw away the old software and write a brand new one from scratch.

Enter "OmniPlanner."

Think of OmniPlanner as a universal translator and a master navigator that can be plugged into any of these robots. It's a single "brain" that understands the core rules of exploration, no matter what kind of body the robot has.

Here is how it works, broken down into simple concepts:

1. The "Universal Brain" (The Planning Kernel)

Imagine a master chef who knows how to cook a perfect meal. Usually, if you want to cook for a vegan, a meat-eater, and a gluten-free person, you need three different recipes.
OmniPlanner is like a chef who has one master recipe (the core logic) but uses different specialized aprons (adaptation layers) depending on who is eating.

For the Drone: The apron tells the brain, "Remember, you can fly in 3D space, but don't hit the ceiling."
For the Robot Dog: The apron says, "You can't fly! Check if the ground is steep before you step, or you'll fall."
For the Submarine: The apron whispers, "Stay close to the walls; don't drift into the open ocean where you can't see anything."

Because the "brain" stays the same, the same software works for all three robots without needing to be rewritten.

2. The Two-Step Dance (Local vs. Global)

When exploring a giant, dark cave, a robot can get confused. It might see a path right in front of it, take it, and end up in a dead-end tunnel.
OmniPlanner uses a two-step dance:

The Local Step (The Microscope): The robot looks at what is immediately around it (within a few meters). It asks, "What can I touch right now? Where can I go without crashing?" It builds a tiny, detailed map of the immediate neighborhood.
The Global Step (The Telescope): The robot zooms out. It looks at the big picture it has built so far. It asks, "Where are the big open spaces I haven't seen yet? Am I stuck in a dead end?"
The Magic: If the robot gets stuck in a dead end (a local problem), the "Telescope" says, "Hey, I see a better path over there! Let's turn around and go to that new branch." This prevents the robot from spinning in circles forever.

3. The Three Superpowers

OmniPlanner isn't just about walking around; it has three specific modes, like a Swiss Army Knife:

The Explorer (Volumetric Exploration):
- Analogy: Imagine walking through a dark house with a flashlight, trying to see every single corner of every room.
- What it does: The robot moves to see as much "unknown space" as possible. It's like a vacuum cleaner trying to suck up all the dust (unknown areas) in the house.
The Inspector (Visual Inspection):
- Analogy: Imagine you are a detective looking at a specific painting on the wall. You don't just walk past it; you walk up close, move your head to see the cracks, and take a high-quality photo.
- What it does: Once the robot finds a structure (like a pipe or a wall), it switches modes. It plans a path to get the perfect angle to take a picture of every part of that object, ensuring nothing is missed.
The Target Runner (Target Reach):
- Analogy: Imagine someone yells, "Go find the red door at the end of the maze!"
- What it does: The robot ignores the need to map everything and focuses purely on getting to that specific spot as fast and safely as possible, even if it has to navigate through a forest of obstacles it has never seen before.

4. Real-World Proof

The researchers didn't just test this in a computer game. They sent these robots into real, scary places:

Underground Mines: Dark, narrow tunnels where a wrong turn means getting stuck.
Forests: Trees everywhere, with uneven ground that would trip up a normal robot.
Submarine Bunkers: Underwater, where visibility is zero and you can't just "fall" if you lose balance.
Ballast Tanks: Tight, industrial metal tanks with narrow holes to squeeze through.

In every single case, the same software guided a flying drone, a walking robot, and a swimming robot to do their jobs perfectly.

Why This Matters

Before OmniPlanner, if you wanted to explore a mine with a drone, you needed one team of engineers. If you wanted to explore the same mine with a robot dog, you needed a different team of engineers to write new code.

OmniPlanner is like a universal remote control. You can point it at a drone, a robot dog, or a submarine, and it just works. This means we can deploy robots faster, cheaper, and more reliably to save lives in disasters, inspect dangerous infrastructure, or explore places humans can't go.

1. Problem Statement

Autonomous robotic systems are increasingly deployed for mapping, monitoring, and inspection in complex, unstructured environments (e.g., underground mines, forests, submarine bunkers). However, existing path planning approaches suffer from significant fragmentation:

Domain Specificity: Most planners are tailored to specific domains (air, land, or sea) or robot classes (e.g., aerial drones vs. ground rovers).
Lack of Generalization: Adapting a planner to a new robot morphology or environment typically requires substantial redesign, reparameterization, or parallel development.
Task Separation: Exploration (mapping unknown space) and Inspection (observing known surfaces) are often treated as distinct problems with limited support for transitioning between them within a unified architecture.

The core challenge is to develop a unified planning framework that can handle volumetric exploration, visual inspection, and target reach behaviors across heterogeneous robotic platforms (aerial, ground, underwater) with minimal retuning.

2. Methodology: OmniPlanner

OmniPlanner addresses these challenges through a domain-agnostic planning kernel decoupled from robot-specific constraints via lightweight adaptation layers.

A. Core Architecture

The framework utilizes a bifurcated local-global planning structure:

Local Planning Module: Constructs a dense, sampling-based graph ( $G_L$ $G_{L}$ ) within a bounded volume around the robot. It handles immediate collision avoidance and local motion feasibility.
- Supports three sampling strategies: Uniform (broad coverage), Gaussian (dense local connectivity), and Hybrid (balance).
- Offers two graph construction modes: Basic (incremental) and Batch (parallel sampling for faster region discovery).
Global Planning Module: Maintains a sparse, persistent graph ( $G_G$ ) that captures the connectivity of the mapped configuration space over time. It enables long-horizon reasoning, repositioning to escape local minima, and "return-to-home" functionality based on endurance limits.

B. Embodiment Adaptation Layer

To support different morphologies, the kernel is specialized through an adaptation layer that instantiates primitives for sampling, collision checking, and edge validation based on the robot's specific constraints ( $C_\mu$ ):

Aerial Robots: Operates in 3D free space; constraints are primarily collision avoidance and kinematic feasibility.
Ground Robots: Leverages a 2.5D elevation map to enforce terrain traversability (slope limits, footing support) and obstacle avoidance.
Underwater Robots: Enforces proximity constraints to observed structures (preventing entry into open water) while navigating 3D space.

C. Modular Behavior Objectives

Three distinct behaviors are instantiated as modular objectives on top of the shared kernel:

Volumetric Exploration (VE): Maximizes information gain by classifying unknown voxels. It uses a global frontier selection mechanism to reposition the robot when local exploration is exhausted.
Visual Inspection (VI): Systematically observes target surfaces. It generates candidate viewpoints, selects a coverage-optimal subset via a greedy strategy, and solves a Traveling Salesman Problem (TSP) to sequence the inspection path.
Target Reach (TR): Guides the robot to a user-defined target in unknown space. It combines global guidance (selecting frontiers that advance toward the target) with local tracking of a "lookahead point" to ensure safety and progress.

3. Key Contributions

Unified Planning Kernel: A single, domain-agnostic architecture that serves as a shared backbone for global and local planning across heterogeneous platforms (multi-rotors, legged robots, differential drives, underwater ROVs).
Modular Behavior Integration: Unifies exploration, inspection, and target reach within one framework, enabling autonomous switching between tasks without external intervention.
Cross-Domain Validation: Extensive validation across aerial, ground, and underwater robots in diverse environments (mines, forests, ballast tanks, submarine bunkers), demonstrating that the same core logic generalizes effectively with only morphology-specific parameter adjustments.
Open Release: The authors commit to releasing the implementation, simulation environments, and field datasets.

4. Results

The framework was validated through extensive simulations and real-world field deployments.

Simulation Studies

Aerial Robots: In large-scale cave environments, OmniPlanner outperformed state-of-the-art baselines (GBPlanner 2.0, ERRT). It achieved 46.9% higher Area Under the Curve (AUC) for exploration volume and was 33.75% faster than ERRT. In confined ballast tanks, it was 48% faster than FUEL.
Ground Robots: In mine simulations, OmniPlanner reduced exploration time by 22.88% compared to TARE and 6.98% compared to DSVP.
Underwater Robots: In submarine crash-site simulations, OmniPlanner achieved 55.56% higher AUC and was 65% faster than the NBVP baseline.
Feature Analysis: The "Batch" graph construction strategy significantly improved region discovery speed compared to incremental methods. Active camera pitch actuation improved inspection coverage from ~62% to ~88%.

Field Experiments

Eight field deployments were conducted across three platforms:

Aerial (AR-1, AR-2): Successfully explored underground mines and ballast tanks, navigating narrow corridors and performing autonomous inspection.
Ground (ANYmal GR-1): Traversed 1.2 km on a university campus and 357 m in a mine, successfully handling uneven terrain and slope constraints.
Underwater (UR-1): Operated in a low-visibility submarine bunker, performing both exploration and inspection missions.
Marsupial Team: A ground robot (carrying an aerial robot) deployed the drone to explore forest areas inaccessible to the ground unit, demonstrating coordinated multi-robot autonomy.

In all cases, the system operated fully onboard with no human intervention, demonstrating robustness in GPS-denied, visually degraded, and cluttered environments.

5. Significance

OmniPlanner represents a paradigm shift from fragmented, domain-specific planning to universal autonomy. By decoupling the planning logic from the robot's physical embodiment, it significantly reduces the engineering overhead required to deploy autonomous systems in new domains. The framework proves that a single planning strategy can effectively generalize across air, land, and sea, enabling scalable, cross-platform robotic solutions for critical applications like search and rescue, infrastructure inspection, and environmental monitoring. Future work aims to extend this to non-holonomic platforms (e.g., fixed-wing aircraft) and integrate perception uncertainty directly into the sampling process.