The research, published in the International Journal of Intelligent Robotics and Applications, tackles one of the most persistent challenges in aerial robotics: enabling small, agile flying machines to find their way through environments that are simultaneously cluttered, only partially visible, and governed by unforgiving physics. Quadrotors must respect strict dynamic constraints while processing high-dimensional sensory input and reacting to complex obstacle configurations, all in real time. Traditional deep reinforcement learning approaches have typically flattened all of this information into simple, unstructured vectors of numbers, discarding the rich spatial and relational structure that human pilots intuitively grasp when weaving through a forest canopy or a debris field.

The team, led by Mohsen Rabdoost Motlagh, working with Mohammad Ali Javadzade and Hossein Hosseini, took a fundamentally different approach. Instead of feeding the drone’s learning algorithm an undifferentiated stream of raw data, they first organize the environment into a dynamic, ego-centric graph. In this representation, each node encodes a local spatial region or an individual obstacle, while the edges connecting those nodes capture traversability relationships — essentially, which routes through the space are physically passable and which are blocked. Crucially, this graph is ego-centric, meaning it is anchored to the drone itself and updated continuously as the vehicle moves, so the representation shifts and reconfigures in step with the robot’s changing viewpoint.

A graph neural network then acts as the system’s perceptual and reasoning engine. GNNs are a class of neural architectures designed specifically to operate on graph-structured data, inspired by the semi-supervised graph convolutional methods pioneered by Thomas Kipf and Max Welling. By passing information along the graph’s edges, the network learns structured, context-aware embeddings — compact mathematical descriptions in which each node’s representation has been enriched by knowledge of its neighbors and their relationships. This allows the drone to reason about obstacles not as isolated points but as part of an interconnected spatial web, capturing configurations and patterns that flat vector inputs simply cannot express.

These graph-derived embeddings are then fused with the quadrotor’s dynamic state — its position, velocity, orientation and other motion variables — and handed to a reinforcement learning agent built on Proximal Policy Optimization. PPO has become one of the most widely used algorithms in continuous-control robotics because of its stability: it constrains each policy update so the agent improves gradually rather than making erratic leaps that can destroy previously learned behavior. In this framework, the PPO agent outputs continuous control commands, steering the drone toward its goal while avoiding collisions, with the graph-based representation supplying a far richer picture of the world than the agent would otherwise possess.

To test the approach, the researchers built their evaluation in PyBullet, a widely used open-source physics simulation environment, and compared their graph-augmented system against standard PPO baselines that rely on flat, unstructured vector inputs — all under identical experimental conditions. The results were striking. In standard scenarios containing eight obstacles, the graph-based framework achieved a success rate of 92 percent, compared with 65 percent for the baseline — a 27 percentage-point improvement. Perhaps even more important for the reliability of the learned policy, the variance in reward across trials dropped by 65 percent, indicating that the system’s behavior is far more consistent and predictable, a critical property for any robot expected to operate safely in the real world.

The most dramatic divergence appeared when the researchers pushed the difficulty up. In high-density environments packed with 28 obstacles, the proposed method maintained an 81 percent success rate while the baseline collapsed to just 38 percent. That gap — a difference of more than double — illustrates precisely where structured relational representations earn their keep. When obstacles multiply and the space between them shrinks, the fine-grained relationships between individual barriers become the dominant factor in determining whether a path exists. A flat vector of sensor readings treats all of that structure as noise; an ego-centric graph preserves it, and the GNN encoder exploits it.

The consistency of the improvements across varying obstacle densities is particularly significant. Many navigation systems perform well in the conditions they were trained on but degrade sharply as complexity increases. By demonstrating robust performance as the environment scales from sparse to dense, the Tehran team has provided evidence that the benefits of graph-based environmental reasoning are not a quirk of a particular scenario but a general property of the representation itself. Structured relational encoding, in other words, enhances not just the performance of learning-based aerial navigation policies but also their robustness, efficiency and scalability.

The work arrives amid a flourishing of interest in combining graph representations with reinforcement learning for robotics. Recent studies have applied graph reinforcement learning to multi-UAV communication coverage, collaborative search and tracking, and decentralized multi-robot navigation, while other teams have pursued deep reinforcement learning for drone navigation in windy conditions, dynamic target tracking, and vision-based obstacle avoidance. What distinguishes the new contribution is its specific fusion of a dynamic, ego-centric environmental graph with a GNN encoder and a PPO continuous-control agent for single-quadrotor navigation, evaluated head-to-head against equivalent flat-input baselines in a controlled simulation setting.

The practical implications reach well beyond the laboratory. Delivery drones, inspection robots, search-and-rescue aircraft and autonomous vehicles operating in disaster zones all face the same fundamental problem: how to move safely through spaces that are cluttered, unfamiliar and only partially observable. A navigation policy that succeeds more than twice as often in dense environments, and that behaves far more consistently from trial to trial, could translate directly into fewer crashes, more reliable missions and lower certification barriers for commercial deployment. The reduced reward variance is especially meaningful in this context, since policy variability is one of the chief obstacles to trusting learned controllers with safety-critical tasks.

There are, of course, caveats. The evaluation was conducted entirely in simulation, and the sim-to-real gap — the often painful transition from virtual physics engines to the messy realities of wind, sensor noise and battery constraints — remains one of the toughest problems in robotics. The authors note that no experimental datasets were generated or analyzed during the study, and the framework’s performance on physical hardware has yet to be demonstrated. Real-world deployment would also require the graph construction and GNN inference to run fast enough to keep pace with the quadrotor’s rapid dynamics, something that modern embedded hardware increasingly makes feasible but that still demands careful engineering.

Nevertheless, the study adds substantial weight to a growing consensus in robotics research: that how we represent the world matters as much as how we learn to act in it. By giving a reinforcement learning agent a structured, relational, self-centered map of its surroundings — one that evolves with every meter flown — the researchers have shown that the drone does not merely react better; it understands better. As autonomous aerial systems are called upon to operate in ever more complex and crowded spaces, techniques that fuse the representational power of graph neural networks with the decision-making strength of deep reinforcement learning look increasingly likely to define the next generation of intelligent machines in flight.

Modern artificial intelligence systems are extraordinarily powerful, but much of their energy and time can be consumed moving data rather than performing calculations. In conventional computing architectures, sensory inputs and intermediate results must travel repeatedly between processors and external memory. This separation is especially costly for neural networks, which perform large numbers of mathematical operations on streams of data and often need to preserve information about what happened moments or even minutes earlier. The human brain takes a different approach. Neurons both process signals and maintain internal states, allowing computation and memory to coexist across densely interconnected networks. FLARE, short for a fully in-memory photonic computing architecture, is designed around a similar principle: sensing, processing and memory are integrated into the same computing fabric rather than treated as isolated stages.

The system combines photonic and electronic mechanisms to create what the researchers describe as reconfigurable photonic neurons. Photonic computing uses light to carry and manipulate information, an approach that can exploit the speed and parallelism of optical signals. Electronic components, meanwhile, can provide storage, control and nonlinear behavior that are difficult to implement efficiently with light alone. In FLARE, these mechanisms are coupled so that the artificial neurons can respond to incoming signals, transform them through a deep nonlinear neural network and preserve aspects of their previous activity. That memory is not a single, uniform function. The architecture supports both long-term and short-term dynamics, allowing it to retain learned or accumulated information while continuing to react quickly to new inputs.

This distinction between memory timescales is central to systems that must operate in changing environments. Short-term dynamics can help a machine interpret rapidly evolving signals, such as motion, acceleration or changing visual information. Long-term retention can preserve a more persistent internal state, allowing the system to use information gathered earlier rather than treating every new signal as an isolated event. The FLARE chip reportedly maintained long-term memory for 7.45 seconds while preserving short-term behavior in the gigahertz range. A gigahertz corresponds to a billion cycles per second, so the result represents a combination of sustained memory and extremely rapid signal dynamics. Rather than forcing a system to choose between storing information and processing it quickly, the architecture is intended to make both behaviors available within the same neural substrate.

At the physical level, the promise of photonic computing comes from the way light can propagate and interact across many channels at once. Optical signals can encode information in properties such as intensity, and integrated photonic circuits can perform transformations as light travels through carefully engineered structures. In a neural network, these transformations can represent the weighted combinations of inputs that form the basis of artificial-neuron computation. The critical challenge is that useful intelligence requires more than fast linear operations. Neural systems must also incorporate nonlinear responses, memory and adaptation. FLARE addresses those demands by combining optical processing with electronic mechanisms, creating neurons whose responses can be reconfigured and whose internal state can evolve over time. The result is not simply an optical accelerator attached to a conventional processor, but an attempt to merge the roles of memory, computation and sensing.

Scale is another important feature of the reported demonstration. The researchers built a monolithic multicore chip with 7,378 neurons, then arranged ten cores into a multilayer FLARE system for the autonomous-navigation experiment. A monolithic design places the relevant elements within a unified chip architecture, a step intended to reduce the bottlenecks that arise when separate components must exchange data. Multiple cores and layers allow the network to expand beyond a small laboratory proof of concept and to process information through a deeper neural structure. The architecture’s reconfigurability is also significant: a system designed to operate in the physical world must adjust to different inputs and tasks rather than relying on a fixed, one-purpose circuit. By integrating memory directly into the neural elements, FLARE is designed to keep intermediate activations close to where they are generated, limiting the costly movement of data through external digital memory.

The researchers tested the architecture in a setting that demands continuous interaction between perception and action: autonomous racing-drone navigation. A racing drone must interpret sensory information, explore its environment and adapt its behavior while moving rapidly through space. Conventional systems often divide these tasks among sensors, processors, memory and control units, creating delays and energy costs as information is repeatedly converted, transferred and processed. In the FLARE demonstration, the multilayer photonic system handled sensing, exploration and adaptation as parts of an integrated neural process. Its reported energy cost of 61.87 attojoules per operation illustrates the potential efficiency of performing computation in place. An attojoule is 10^-18 joules, an extraordinarily small unit of energy. The figure is reported at the system level, making it relevant to the complete demonstrated architecture rather than only to an isolated optical operation.

The most striking implication is that FLARE could help shift artificial intelligence away from the traditional sequence of “sense, store, compute and act.” In embodied systems such as drones, robots and autonomous vehicles, intelligence is inseparable from the physical world. Sensors generate continuous streams of data, and useful decisions depend on temporal context: what the machine detected earlier, how the environment is changing and whether a previous action produced the expected result. A network with integrated memory can preserve this context as part of its ongoing activity. A network that also processes signals photonicly could, in principle, handle high data rates without sending every intermediate result to a distant electronic memory. The combination may be particularly valuable where size, energy consumption and response time are tightly constrained.

The work nevertheless represents a pathway rather than a finished replacement for conventional artificial intelligence hardware. The reported results establish the architecture’s neuron count, memory retention, high-speed dynamics and autonomous-navigation demonstration, but practical deployment will depend on how such systems perform across a broader range of tasks and operating conditions. Photonic and electronic components must remain precisely coordinated, and large integrated neural systems must be manufactured, programmed and calibrated reliably. Memory that lasts several seconds may be useful for navigation, but different applications could require other timescales. Likewise, energy per operation is only one part of a system’s total cost; sensing, communication, control and training procedures also matter. These are engineering questions that will shape whether fully in-memory photonic systems can move from specialized demonstrations into everyday machines.

FLARE’s significance lies in the way it brings several long-standing ambitions together on one platform. It uses light for fast, parallel signal processing, electronics for memory and control, neural-network organization for nonlinear computation, and integrated architecture for direct interaction with sensory inputs. The resulting chip does not merely imitate the brain’s appearance; it targets a functional property that makes biological intelligence efficient: memory and computation are deeply intertwined. With thousands of photonic neurons, seconds-long retention, gigahertz-scale short-term dynamics and a low reported energy cost in drone navigation, the system offers a glimpse of machines that compute less by moving data and more by transforming information where it already exists. If the approach continues to scale, future autonomous devices could become faster, more adaptive and substantially more energy-efficient without relying on the memory bottlenecks that define much of today’s AI hardware.

The research addresses a problem known as informative path planning. Unlike ordinary navigation, in which a vehicle travels from one known location to another, informative path planning requires a robot to choose a route that will collect valuable information. An unmanned aerial vehicle searching for a missing person, for example, must decide which regions to photograph, how to arrange those observations, and when the evidence gathered so far justifies changing direction. The vehicle is not merely moving through space; it is selecting a sequence of measurements that may reduce uncertainty about the world. Every decision therefore has two consequences: it determines where the robot will travel and it influences what the robot will know when it arrives.

This challenge is becoming more urgent as unmanned vehicles spread into military surveillance, package delivery, scientific exploration, precision agriculture, infrastructure inspection, and emergency response. Market research company Grand View Research estimates that the unmanned systems market was worth $29.3 billion in 2025 and could reach $67.6 billion by 2033. As these systems operate in more complex environments, they must cope with incomplete maps, changing conditions, limited battery capacity, communication delays, and the high cost of transmitting or processing data. A drone surveying a forest may have only a rough estimate of where a person is located. Images collected during the first part of the mission can alter that estimate, but using every new image to immediately calculate a completely new route may consume time and energy that the robot cannot afford to waste.

At one end of the planning spectrum is a fully adaptive strategy. In this model, the vehicle follows a route for a short period, receives new observations, and then recomputes its next move whenever information becomes available. If a drone’s camera detects a possible signal, an unusual shape, or an area that appears more promising than expected, the system can react immediately. This flexibility can produce highly effective routes because the vehicle continually responds to evidence. However, the computational burden can be substantial. Replanning requires algorithms to evaluate possible future paths, compare their expected information value, account for travel costs, and often communicate new instructions to the vehicle. In a rapidly changing mission, repeated calculations may become a bottleneck rather than an advantage.

At the opposite extreme is a nonadaptive strategy. The operator or planning system calculates a route before the mission begins, and the vehicle follows it regardless of what it discovers. This approach is easier to execute and may be faster because it avoids repeated optimization. It can also reduce communication demands and make battery consumption easier to estimate. Its weakness is that it treats the future as if it were already known. A search drone could continue scanning low-probability areas even after early observations indicate that the missing person is likely elsewhere. A utility-inspection robot might persist along a predetermined sequence after detecting evidence that the fault lies in another direction. The result may be operationally simple but scientifically and practically inefficient.

Ghuge, working with Rayen Tan and Viswanath Nagarajan of the University of Michigan, investigated a middle path between these two extremes. Their proposed framework limits adaptivity by dividing a mission into a small number of sequential rounds. During the first round, the robot follows a designated route without changing it in response to observations collected along the way. At the end of that round, the system uses the accumulated information to recompute the next route. The process can then be repeated, allowing later stages of the mission to become increasingly responsive without requiring continuous replanning. Technically, the method separates data collection from major optimization decisions. Instead of solving a new path-planning problem after every observation, the system solves it only at selected checkpoints.

The researchers tested this limited-adaptivity strategy through computerized simulations and compared it with a fully adaptive model. Their results indicate that two adaptive rounds could be about 15 times faster than a continuously adaptive approach. That acceleration came with a relatively small financial or operational penalty: the cost of the two-round strategy was only 12% higher than that of a comparable fully adaptive search. The simulations also showed that additional rounds produced diminishing returns. After the first three rounds, accuracy did not increase markedly, suggesting that constant route revision may deliver little extra value once the system has incorporated the most important early information. In practical terms, the robot may need only a few opportunities to reconsider its mission rather than an unbroken stream of decisions.

The reason this compromise works is that early observations often provide the largest improvements in situational understanding. Before a mission begins, the robot may face broad uncertainty about the location of a target or the source of a problem. The first sweep can eliminate large portions of the search area or reveal patterns that change the probability distribution over possible locations. A second route can then focus on the most promising regions. By the third round, the remaining uncertainty may be narrower, while the cost of further optimization continues to accumulate. Limited adaptivity attempts to capture the high-value information gained early in a mission while avoiding the computational and logistical expense of reacting to every minor change in the data.

The implications extend beyond search-and-rescue operations. A drone inspecting power lines could fly an initial route, analyze signs of damage, and then direct a second pass toward the structures most likely to contain the source of an outage. In agriculture, an autonomous vehicle could survey a field, identify areas with unusual moisture or plant stress, and allocate later measurements more intelligently. Scientific robots exploring remote terrain could use staged planning to balance coverage with the need to investigate anomalies. In each case, the system must optimize more than geographic distance. It must weigh the expected value of information against flight time, battery use, sensor operation, data-transfer requirements, and the consequences of delaying a response.

The study does not suggest that continuous adaptation is never useful. In missions involving rapidly changing hazards, extremely valuable targets, or sudden safety threats, immediate replanning may justify its computational cost. Instead, the findings indicate that adaptivity should be treated as a limited resource and scheduled strategically. A robot that changes its solution only two or three times may retain most of the performance advantage of a fully adaptive system while operating much faster and with fewer communication demands. For organizations deploying autonomous vehicles, that balance could determine whether a system remains theoretical or becomes practical. In a winter search for a missing hiker, minutes matter; a route that is slightly less precise but can be executed far sooner may ultimately offer the better chance of success. The research, published as “Informative Path Planning with Limited Adaptivity” in INFORMS Journal on Computing, provides a mathematical and computational foundation for making that trade-off deliberately rather than assuming that more frequent decision-making is always better.

Subject of Research: Limited-adaptivity algorithms for informative path planning by autonomous vehicles

Article Title: Informative Path Planning with Limited Adaptivity

News Publication Date: 27-May-2026

Web References: https://www.mccombs.utexas.edu/faculty-and-research/faculty-directory/profile/?username=rg53727 ; https://www.grandviewresearch.com/industry-analysis/unmanned-systems-market-report ; https://pubsonline.informs.org/doi/abs/10.1287/ijoc.2024.0893

References: INFORMS Journal on Computing, DOI: 10.1287/ijoc.2024.0893

Keywords: Computer science, algorithms, applied mathematics, business, autonomous vehicles, drones, informative path planning, adaptive robotics, search and rescue, unmanned systems