⚡ electrical engineering

SEAMLiS: Visibility-Aware Safety for Perception-Limited Multi-Robot Exploration

This paper introduces SEAMLiS, a modular execution-layer safety framework for decentralized multi-robot exploration that ensures collision-free navigation under limited sensing by integrating a visibility-aware attitude filter with a Control Barrier Function-based positional filter, thereby preserving the efficiency of upstream exploration strategies while preventing late obstacle detection.

Original authors: Taekyung Kim, Rahul H Kumar, Aswin D. Menon, Tzu-Hsiang Lin, Dimitra Panagou

Published 2026-07-14

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Taekyung Kim, Rahul H Kumar, Aswin D. Menon, Tzu-Hsiang Lin, Dimitra Panagou

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 team of brave, tiny robot explorers sent into a pitch-black cave they've never seen before. Their mission? To map every inch of the darkness as fast as possible. They have a super-smart brain (the "planner") that tells them where to go next, and a set of eyes (the "sensor") that can only see a short distance ahead and in a narrow cone, like a flashlight beam.

Here's the problem: In the past, these robots were told to be "information greedy." This meant their brains would constantly swivel their flashlight beams to look at the most interesting, unexplored corners of the cave, even if that meant turning their backs on the path they were walking.

The "Optimistic" Trap
The paper argues that this strategy is dangerous. Imagine a robot walking forward while spinning its head to look at a cool rock formation to its left. It's so focused on the rock that it doesn't notice a giant boulder directly in front of it until it's too late. Because the robot can only stop so quickly (it has "bounded actuation"), by the time it sees the boulder, it's already too close to stop. The result? A crash.

The authors explicitly rule out the idea that simply having a "safety controller" at the end of the line can fix this. They show that if the robot's brain plans a path through "invisible" space and the robot's eyes are looking the wrong way, the safety controller gets overwhelmed and can't stop the crash in time.

Enter SEAMLiS: The "Gatekeeper" Guardian
To fix this, the researchers created a new system called SEAMLiS (Safe Exploration for Autonomous Multi-Robot Systems Under Limited Sensing). Think of SEAMLiS not as a new brain, but as a strict Gatekeeper standing between the robot's brain and its body.

The Gatekeeper has a simple job: It checks if the robot is looking in the right direction before it lets the robot move.

The "Go" Signal: If the robot is walking toward a safe, known area, the Gatekeeper says, "Go ahead! Spin your head and look at that cool rock!" This keeps the exploration fast and efficient.
The "Stop and Look" Signal: But the moment the robot approaches the edge of what it knows—the boundary between the safe, mapped floor and the dark, unknown cave—the Gatekeeper intervenes. It says, "Hold on! You can't look at the rock right now. You need to point your flashlight straight ahead to make sure there are no boulders in your path."

If the robot tries to keep looking at the rock, the Gatekeeper overrides the command and forces the robot to look straight ahead (a "velocity-tracking" mode) until it's safe again.

How They Tested It
The team didn't just guess this would work; they put it to the test in three different ways:

Randomized Simulations: They ran 100 different computer trials with 1, 2, or 3 robots in messy, obstacle-filled environments. In these tests, the old "look-at-the-rock" strategies crashed repeatedly (sometimes 100% of the time with 3 robots), while SEAMLiS achieved a 100% success rate with zero collisions.
Isaac Sim: They tested it in a high-fidelity physics simulator that mimics real-world gravity and movement. The results were the same: the old methods crashed, SEAMLiS didn't.
Real Hardware: They flew two actual Crazyflie quadrotors (tiny drones) in a room about 3.35 m by 3.50 m. They set the robots to fly at a speed of 1.35 m/s with a sensor range of just 0.25 m. In this real-world test, the robot using the old "constant yaw-rate" (spinning its head) crashed into an invisible obstacle. The robot using SEAMLiS, however, successfully avoided the obstacle by keeping its eyes on the path.

The Bottom Line
The paper proves that you don't have to choose between being fast and being safe. By adding this "Gatekeeper" layer, robots can still explore efficiently, but they are forced to keep their eyes on the road when they get close to the unknown.

The results show that SEAMLiS eliminates the hidden-obstacle collisions that plague other methods while keeping the exploration time almost as fast as the risky, unguarded methods. It's a safety net that catches the robot before it falls, ensuring that even in the dark, the team can map the cave without getting hurt.

Technical Summary: SEAMLiS

Problem Statement
Autonomous multi-robot exploration in unknown environments typically relies on an efficiency-oriented stack where goal selection (e.g., frontier-based) and local planning treat unobserved space as collision-free to maximize information gain. However, under finite sensing ranges and limited fields of view (FoV), this separation of planning and safety creates a critical vulnerability. When a robot's yaw (sensor heading) is controlled to maximize information gain rather than to observe the direction of motion, the robot may translate toward hidden obstacles that lie outside its current FoV. If these obstacles are detected only after the remaining stopping distance becomes insufficient for the robot's bounded actuation, collision avoidance becomes infeasible, even if a safety controller (such as a Control Barrier Function) is active. This failure mode is particularly acute in decentralized settings where robots do not share full occupancy maps, meaning a robot cannot rely on other agents' observations to avoid locally hidden hazards.

Methodology
The paper proposes SEAMLiS (Safe Exploration for Autonomous Multi-Robot Systems Under Limited Sensing), a modular execution-layer safety framework designed to plug into existing decentralized exploration stacks without replacing upstream goal allocators or planners. SEAMLiS operates by enforcing two distinct safety filters:

Visibility-Aware Attitude Filter (Gatekeeper):
- Objective: To ensure the robot maintains visibility of the critical boundary between its known-free region and the unknown region early enough to preserve a braking margin.
- Mechanism: The filter employs a "gatekeeper" algorithm that evaluates a nominal visibility-promoting yaw policy (which greedily maximizes information gain) against a backup velocity-tracking yaw policy (which aligns the sensor with the direction of motion).
- Logic: It defines a "visibility-safe set" based on a visibility constraint function, $h_{vis}$ . This function calculates whether the robot has sufficient path distance to rotate the critical boundary point into its FoV while retaining a worst-case braking distance ( $d_{br}$ ).
- Execution: If the nominal policy can be executed for a certified duration while keeping the system within the visibility-safe set, the gatekeeper allows the information-greedy yaw. If not, it immediately switches to the velocity-tracking backup policy to ensure the motion-relevant unknown boundary remains observable.
CBF-Based Positional Filter:
- Objective: To avoid collisions with known obstacles, newly detected obstacles, and neighboring robots.
- Mechanism: This filter uses High-Order Control Barrier Functions (HOCBFs) to modify the nominal acceleration command.
- Constraints: It enforces constraints for:
  - Obstacle Avoidance: Ensuring distance to known and newly detected obstacles remains above a safety radius.
  - Inter-Agent Separation: Ensuring distance to neighboring robots remains above a safety radius, utilizing a decentralized pairwise allocation of responsibility.
- Integration: The positional filter acts as a quadratic program (QP) that minimally modifies the nominal path-tracking command to satisfy these constraints.

Key Contributions

Formalization of a Safety Gap: The paper identifies and formalizes the specific safety failure mode where optimistic planning combined with visibility-maximizing yaw control delays hidden-obstacle detection until safe avoidance is infeasible under bounded actuation.
Modular Safety Framework: SEAMLiS is designed as a plug-and-play execution layer that augments arbitrary, potentially asynchronous, multi-robot goal assignment and planning modules, preserving the scalability of existing decentralized pipelines.
Visibility-Aware Attitude Filter: The design of a gatekeeper-style mechanism that dynamically switches between information-greedy and motion-aligned yaw policies, ensuring that critical unknown boundaries are observed with sufficient braking margin.
Theoretical Guarantees: The authors provide sufficient conditions (Assumptions 1 and 3) under which the combined attitude and positional filters guarantee collision avoidance for all time, even in the presence of initially unknown obstacles, provided the sensing range is sufficient to detect obstacles before the braking distance is exhausted.
Validation: The framework is validated through randomized simulations, Isaac Sim environments, and hardware experiments with two Crazyflie quadrotors.

Results

Simulation: In randomized simulations with 1, 2, and 3 robots using both frontier-based and decentralized CoScan-inspired goal allocators, SEAMLiS achieved a 100% success rate with zero collisions across all tested settings. In contrast, baseline policies (constant yaw-rate and pure visibility-promoting) exhibited collision rates up to 1.00 in multi-robot scenarios.
Efficiency: SEAMLiS retained much of the exploration efficiency of the visibility-promoting baseline. While the velocity-tracking baseline was also collision-free, it was overly conservative and failed to complete exploration tasks within the time limits. SEAMLiS completed tasks in times comparable to the efficiency-oriented baselines.
Hardware: In physical experiments with Crazyflie quadrotors, the constant yaw-rate baseline resulted in a collision with an initially unknown obstacle that was not shared in the local map. SEAMLiS successfully avoided the collision by switching to the velocity-tracking backup policy when the visibility margin was threatened.

Significance
The paper claims that SEAMLiS addresses a critical gap in decentralized multi-robot exploration: the lack of formal safety guarantees when sensing is finite-range and FoV-limited. By decoupling the efficiency of exploration planning from the enforcement of safety, SEAMLiS allows robots to maintain high exploration speeds while rigorously ensuring that hidden obstacles are detected early enough to allow for bounded-actuation avoidance. The framework demonstrates that safety does not require sacrificing the efficiency of information-greedy strategies, provided that the execution layer actively manages the visibility of critical unknown boundaries.

Technical Summary: SEAMLiS

More like this