The Shape-Shifting Drone That Squeezes Through Gaps a Human Hand Can Barely Fit

Imagine a collapsed building, smoke still rising, a survivor somewhere inside. The rescue drone you send in is agile, but the only route to the victim is a ragged gap between two slabs of concrete — barely wider than the drone's body. The drone's navigation software, however, sees that gap and refuses to enter. Not because the gap is too small. Because the math says the cost of being near two walls simultaneously is simply too high.

This is not a hypothetical failure mode. It is a well-documented limitation in how most autonomous robots currently decide where to go. And it is exactly the problem that Harsh Modi, Xiao Liang, and Minghui Zheng set out to fix in their 2026 paper, Constrained MPC-Based Motion Planning for Morphing Quadrotors in Ultra-Narrow Passages under Limited Perception (Modi et al., 2026).

Their solution is elegant in the way good engineering often is: not more computation, not a more powerful sensor, but a rethink of a mathematical cost function that most researchers had quietly accepted as given. The result is a framework that lets a shape-shifting quadrotor — a drone that can fold its arms inward like a bird tucking its wings — plan its own body geometry and its flight path simultaneously, threading through gaps that would stump a conventional planner entirely.

The Science

To understand what the team built, you need to understand three interlocking ideas: morphing drones, model predictive control, and the obstacle-avoidance cost problem.

Morphing quadrotors are drones whose arm geometry can change mid-flight. By folding or retracting their arms, they can reduce their effective width, allowing them to fit through openings that a rigid-frame drone could not. The catch is that changing the arm configuration changes the drone's aerodynamics, its center of mass, and its moment of inertia — all the physical properties that determine how it flies. A good planner must account for all of that simultaneously.

Model Predictive Control — MPC — is the planning approach the team uses. Think of it as a rolling chess game: at every timestep, the controller looks a fixed number of steps ahead (the "prediction horizon"), simulates many possible futures using a model of the drone's physics, and picks the action sequence that minimizes a cost function. Then it executes the first step, looks ahead again, and repeats. MPC is popular in robotics and aerospace because it can enforce hard constraints (don't exceed this speed, don't collide with that wall) while optimizing performance. The team uses the acados solver — an open-source, numerically efficient MPC framework — to run the optimization fast enough for real-time flight.

The cost function is where the innovation lives. In MPC, you guide the robot toward desirable behavior by assigning numerical costs to states you want to avoid. The standard approach for obstacle avoidance is the artificial potential field (APF): model every obstacle as a repulsive "hill" in cost space, so the optimizer naturally steers the robot away. The problem, as Modi et al. (2026) identify clearly, is that when two obstacles are close together — forming a narrow corridor — their repulsive hills add up. The combined cost at the center of the corridor can be higher than the cost of not entering at all. The planner, doing its job correctly, concludes that going through the gap is too expensive and looks for another route. If there is no other route, it simply stops.

The team's fix is a smooth exponential obstacle cost that behaves differently inside a narrow passage. Rather than summing raw repulsive potentials, the new formulation introduces a cost reduction factor that detects when the robot is surrounded by obstacles on multiple sides — the signature of a narrow corridor — and suppresses the aggregate cost in that region. Near any single obstacle surface, the cost remains steep: collisions are still punished hard. But at the center of a navigable gap, the cost stays low enough that the optimizer recognizes the path as traversable.

Crucially, the function is smooth and differentiable everywhere, which matters enormously for MPC. Hard thresholds or discontinuous activation functions break the gradient-based solvers that MPC relies on. The exponential form avoids this entirely.

The sensor feeding the planner is a 2D LiDAR — a rotating laser that emits pulses in a flat horizontal plane and measures the distance to whatever reflects each pulse. The result is a 360-degree scan of distances to nearby surfaces, updated many times per second. LiDAR is well-established in autonomous vehicles and robots; what is notable here is how the team uses it. Rather than first building a map of the environment and then planning through that map (the standard pipeline), they feed the raw LiDAR point cloud directly into the MPC cost function. Each detected point becomes an obstacle input to the cost computation. This sidesteps an entire processing stage and lets the drone react to obstacles of arbitrary shape — no assumptions about walls being flat or gaps being rectangular.

What They Found

The paper presents results across two regimes: simulation and physical hardware experiments. In simulation, the team constructed scenarios involving narrow corridors and gaps that would be representative of the tightest real-world passages a drone might encounter — think a gap between two structural pillars, or a doorframe in a partially collapsed structure.

The baseline comparison is direct: run the same scenario with a classical APF-based obstacle cost versus the new exponential cost. The APF planner, as expected from the theory, fails to enter the narrowest gaps. The cost landscape it perceives treats the corridor entrance as a barrier rather than a passage. The drone either stops at the threshold or finds a path around the obstacles — and when no such path exists, it stalls.

The exponential cost planner navigates through. The drone's planned trajectory passes through the center of the gap, maintaining clearance from both walls, and emerges on the other side. The morphology planner concurrently commands the arm-folding geometry that minimizes the drone's width at the moment of traversal, coordinating body shape with the flight path in the same optimization loop.

The physical experiments — conducted on actual morphing quadrotor hardware — replicate this behavior in the real world, confirming that the simulation results are not artifacts of idealized physics. The drone successfully traversed narrow corridors where the conventional approach failed, using only its 2D LiDAR for environmental perception.

Traversal Success: Exponential Cost vs. APF Cost

Label	Value
Exponential Cost (Proposed)	1 success (1=pass, 0=fail)
APF Cost (Baseline)	0 success (1=pass, 0=fail)

Binary outcome reported in simulation and hardware experiments (Modi et al., 2026). APF = Artificial Potential Field.

From a computational standpoint, the acados-based implementation runs efficiently enough for real-time control — a non-trivial requirement given that the MPC must solve a constrained nonlinear optimization problem at every control step. The direct LiDAR integration, while adding sensor data to the cost computation, does not require the full mapping pipeline that would otherwise dominate processing time.

Framework Capability Comparison: Proposed vs. Classical MPC

Label	Value
Narrow-Gap Traversal	9
Arbitrary Obstacle Shapes	9
Real-Time Computation	8
Smooth Cost Landscape	9
No Hard Thresholds	9

Qualitative ratings (1–10) derived from comparative discussion in Modi et al. (2026). Not directly measured numeric scores.

Why This Changes Things

The immediate application the authors target is search-and-rescue and inspection in cluttered, confined environments. These are exactly the scenarios where today's drones most conspicuously fail. A drone inspecting the inside of a bridge pier, a pipeline, or a ventilation shaft must navigate passages that were never designed with robot access in mind. A rescue drone searching a disaster site faces gaps and corridors that are irregular, unpredictable, and unavoidable.

The current workaround is to either use smaller, fixed-frame drones (which carry less payload and are less stable) or to rely on human teleoperation (which is slow, fatiguing, and requires line-of-sight expertise). A morphing quadrotor that can plan its own geometry and navigate autonomously through tight spaces addresses both limitations simultaneously.

But the paper's reach extends further than drones. The authors explicitly frame the exponential cost function as robot-agnostic — applicable to any mobile robot using MPC for navigation. Ground robots navigating building interiors, surgical robots threading through anatomical passages, underwater vehicles inspecting submerged infrastructure: any platform where the classical APF approach causes the planner to treat navigable tight spaces as impassable barriers could benefit from this formulation. That is a wide category.

There is also something worth pausing on about the nature of the fix. The team did not solve this problem by adding sensors, increasing compute, or making the drone bigger and more capable. They solved it by identifying a conceptual flaw in a widely used mathematical formulation and replacing it with something more faithful to physical reality. Narrow corridors are not inherently dangerous — they are just geometrically constrained. A cost function should reflect that distinction. The insight is simple in retrospect, which is often the sign of a genuine contribution.

Key System Design Choices in the Proposed Framework

Label	Value
Prediction Horizon Stages	1 MPC rolling window
Morphology + Trajectory Co-planning	1 MPC rolling window
Map-Free LiDAR Integration	1 MPC rolling window
Smooth Exponential Cost	1 MPC rolling window
Cost Reduction Factor	1 MPC rolling window

Architectural summary from Modi et al. (2026). All five components are active in the proposed system.

The direct LiDAR-to-MPC pipeline is separately significant. The conventional approach to autonomous navigation — sense, then map, then plan — introduces latency and accumulates errors at every stage. By skipping the intermediate map, the team's framework reacts to sensor data faster and makes no assumptions about obstacle geometry. This is especially valuable in dynamic environments, where obstacles may shift between when the map was built and when the drone acts on it.

The acados solver underpinning the MPC is an open-source tool, and the team has released their full implementation code on GitHub. That combination — open solver, open code, general-purpose cost function — means other researchers can directly extend this work to their own platforms without starting from scratch.

What's Next

The paper is candid about its current scope and what remains to be done. The 2D LiDAR constrains perception to a single horizontal plane. Real environments are three-dimensional: a gap might be navigable in the horizontal scan but obstructed above or below. Extending the framework to 3D LiDAR or depth cameras is the natural next step, though it also multiplies the computational load of feeding sensor points directly into the optimization.

The morphing model used in the current work captures the key physics of arm-folding but is necessarily a simplification. Real morphing hardware introduces mechanical compliance, actuation delays, and aerodynamic effects that grow more complex as the drone speeds up or operates close to surfaces — a phenomenon called ground effect and wall effect that alters lift unpredictably. More detailed aerodynamic modeling inside the MPC would improve fidelity at the cost of solver complexity.

There is also the question of dynamic obstacles. The current framework assumes the environment is static — walls and pillars, not moving people or machinery. Extending the cost function and the prediction horizon to account for moving obstacles is a substantial open problem, though the direct LiDAR integration does at least mean the drone would see a moving obstacle quickly rather than relying on a stale map.

Finally, the hardware experiments demonstrate feasibility but not yet robustness at scale. Flight through a single narrow corridor under controlled laboratory conditions is a meaningful proof of concept. Autonomous navigation through a sequence of unpredictable gaps in a real disaster environment is a harder bar — one that will require extensive field testing, fault tolerance engineering, and probably integration with higher-level mission planners.

None of these are criticisms so much as natural coordinates for the next several years of work. Modi, Liang, and Zheng have identified and solved a foundational problem that was quietly limiting an entire class of robotic navigation systems. The exponential cost function they propose is simple enough to explain in a single paragraph, general enough to apply across dozens of platforms, and — unlike so much of what gets published in robotics — immediately demonstrable on real hardware.

When a rescue drone one day threads through rubble to find a survivor, part of the math making that possible might trace back to the insight that a narrow gap is not, by definition, a wall. It just looks like one to the wrong cost function.