Two New Papers Show Drones Learning to Navigate Without GPS. Here's What They Actually Demonstrate.

This isn't entirely new. Relative navigation for tracking has been demonstrated before, where you anchor your planning and control to a visible target object. But those systems break down the moment you lose sight of the target, which happens constantly in real search operations. You're flying through a forest, you spot a person, you lose them behind trees, now what? Previous relative tracking systems basically had no answer except "try to reacquire."

What HUNT adds, and this is the genuinely novel contribution, is a unified framework that handles both the search phase (when no target is visible) and the tracking phase (when one is) within the same relative formulation. During search, the drone navigates reactively using only instantaneous observables. Once a target appears, the same perception-control pipeline transitions to tracking without any mode switching or handoff between different subsystems. The paper demonstrates this in outdoor experiments in dense forests, container compounds, and simulated search-and-rescue scenarios with vehicles and mannequins.

Now, I should note some limitations. The outdoor experiments, while impressive, represent a fairly narrow slice of the possible failure modes. Dense forest is hard, but it's also relatively uniform. The paper doesn't address, at least not in detail, what happens in environments with highly variable structure or dynamic obstacles beyond the target itself. And the "search" behavior, while functional, appears to be fairly simple reactive traversal rather than any kind of intelligent search strategy. The drone isn't reasoning about where a target might be; it's just flying through the environment while looking for one. That's a reasonable starting point, but it's worth being clear about what's being claimed.

The second paper, AirDreamer, takes a completely different approach that's rooted in the world models paradigm that's been gaining traction in reinforcement learning research over the past few years. The basic idea is that instead of learning a direct mapping from observations to actions (which tends to produce brittle policies that fail in new environments), you first learn a model of how the world works, then use that model to plan actions.

The authors frame this as "inspired by animal navigation behavior," which is the kind of claim that makes me reach for the methodology section. What they actually mean, as far as I can tell, is that animals don't navigate by following predefined rules but by building internal models of their environment and using those models to predict the consequences of their actions. Whether this is actually how animal navigation works is a question for neuroscientists, but the computational framework is well-defined: learn a world model from experience, then use reinforcement learning to train a policy on top of that model.

What's interesting about AirDreamer specifically is the sparse reward function. Most RL-based navigation systems use heavily shaped rewards, basically giving the robot a dense signal about how well it's doing at every timestep. This makes training easier but often produces policies that exploit the reward shaping rather than actually solving the navigation problem. You end up with drones that are really good at maximizing whatever proxy metric you designed rather than actually getting to the goal. AirDreamer uses a sparse reward (you get reward when you reach the target, nothing otherwise) and claims this avoids local minima traps and produces more robust behavior.

The paper reports a 5.3% higher navigation success rate than the best baseline in "challenging maps," which, I know I'm being picky here, but that's a pretty modest improvement and the phrase "challenging maps" is doing a lot of work. What makes a map challenging? How were the baselines selected? The paper claims effective sim-to-real transfer "without any tuning during deployment," which would be remarkable if true, but the details of the real-world experiments aren't fully clear from the abstract.

It's worth noting that these two approaches aren't necessarily competing. HUNT is solving a specific operational problem (unified search and track) with a specific technical approach (relative navigation). AirDreamer is exploring whether world models can produce more generalizable navigation policies. You could imagine a system that uses world-model-based planning for high-level decisions and relative navigation for low-level reactive control. Whether that would actually work better than either approach alone is an open question.

The broader context here is that autonomous drone navigation in unstructured environments remains genuinely hard. GPS-denied flight, high-speed obstacle avoidance, and generalization to new environments are all active research problems. Both of these papers make incremental contributions to that larger effort. Neither is a breakthrough that solves the problem. Both are worth reading if you're working in this space.

What I'd want to see next from both groups is more rigorous comparative evaluation. HUNT needs to be tested against a wider range of baselines and in more diverse environments. AirDreamer needs clearer documentation of its real-world experiments and a more detailed analysis of where and why it fails. Both papers would benefit from longer-duration experiments that stress-test the systems over hours rather than minutes.

The tendency to treat every new paper as a revolution obscures the actual structure of scientific progress. These are two solid pieces of work that advance the state of the art in specific, well-defined ways. That's how research actually works. It's less exciting than "AI drones that can fly themselves," but it's more useful for understanding what's actually happening in the field.

Sarah Williams · Yesterday · 4 min