Imagine teaching someone to read a map. If they only ever see it pinned to a wall the same way up, they may learn the route by its shape — “turn at the big square, then the long corridor.” Spin the map a quarter turn and they are lost, even though it is the same map. What you actually wanted was for them to read any orientation: to understand the layout, not memorize one picture of it.
This is the core idea behind data augmentation in reinforcement learning. When we train a drone’s navigation model on a fixed set of maps, the risk is that it memorizes those specific layouts instead of learning a general skill. Augmentation fights this by showing the model many transformed versions of each map — rotated, mirrored, perturbed — so it cannot lean on any single orientation. The hope: a model that covers an unseen space well, regardless of how that space happens to be turned.
That is the textbook story. Our own experiments told a more interesting one.
The promise: orientation invariance
We built a training environment that, on every reset, randomly rotates each map by quarter turns and mirrors it horizontally and vertically. From a handful of base maps, this produced a large pool of distinct training views. Everything else stayed fixed — same PPO setup, same maps, only the augmentation changed, so we could read the effect cleanly in a controlled comparison.
The motivation was sharp. A previous model had developed a strong, one-sided turning habit: it relied heavily on a single turn direction. That could mean one of two things. Either it had discovered a genuinely smart traversal strategy, or it had overfit to the particular orientations of the training maps. Augmentation is the classic way to tell the difference. If the habit was real and transferable, coverage should hold. If it was memorization, the habit should dissolve — and we would at least be left with an honest, orientation-invariant model.
The habit dissolved. The model stopped favoring one turn direction and spread its behavior across mirrored turns, sideways moves, and more frequent sensor scans. That confirmed the habit had been overfitting. But coverage also dropped by a few points. The lesson from dev-log 20 was uncomfortable: an orientation-invariant model can be worse than a specialized one when the training and test maps come from the same kind of environment. The classic augmentation results assume training and test data look genuinely different. Ours did not — our maps are semantically similar to one another — so augmentation added more noise than value.
The twist: less augmentation did worse
The obvious next thought was “the mirrors are the problem.” A mirrored corridor opens from the opposite side; maybe that introduced a kind of semantic noise. Rotations, by contrast, are clean — a rotated map is the same map seen from a different angle. So we tried rotations only, no mirrors, in another controlled comparison.
The result, in dev-log 22, inverted intuition: less augmentation did worse, not better. Rotations alone (four orientations) regressed coverage more than the full set (rotations plus mirrors, sixteen orientations). Removing the mirrors did not clean things up — it took something away.
Why? The full set, with mirrors, gives the model symmetry axes to organize around. When a layout and its mirror image both appear, the model can learn one symmetric strategy that covers both — a compact, reusable representation. Strip the mirrors out and those axes vanish. The model still has to learn rotation invariance across four orientations, but now it has no symmetric structure to compress its knowledge around. It settles into a middling, undertrained policy. Crucially, it also failed to find a compensation strategy: where the full set learned to scan and strafe more to make up for the lost turning habit, the rotations-only model just stalled.
The principle: augment with the task’s real symmetries
Put the two findings together and a clean rule emerges. Augmentation only helps when it reflects a symmetry the task genuinely has — and applying a partial or wrong symmetry can do real damage.
A coverage task on these maps is rotationally symmetric: a path that covers a space is just as valid turned ninety degrees. The mirror symmetry, in our setup, turned out to be load-bearing too — it gave the model the structure it needed to generalize. Half-measures broke that structure. And in tasks where left and right genuinely differ — a rule that depends on handedness, a layout where one side matters — mirroring would corrupt the very signal the model needs. The mirror flips left into right, and if the task cares about that distinction, you have taught the model something false.
The takeaway is not “always augment” or “never augment.” It is: identify the symmetries your task actually possesses, then augment with exactly those — fully, not partially.
Augmentation do’s and don’ts
Do:
- Match augmentation to the symmetries the task truly has (rotational, mirror, or neither).
- Apply a symmetry fully if you apply it at all — a complete set of orientations beats a partial one.
- Use augmentation as an honesty test: if a learned habit dissolves under it, that habit was probably overfitting.
- Expect augmentation to pay off most when training and test data look genuinely different.
Don’t:
- Mirror a task that depends on left/right asymmetry — you will break the signal the model relies on.
- Assume “less augmentation is safer.” A partial symmetry set can underperform both the full set and no augmentation at all.
- Port a technique from the literature without checking that its underlying assumption fits your setup.
- Treat a dissolved habit as automatically good news — measure whether coverage held up too.
Closing
Augmentation is one of the most reliable generalization tools in machine learning, but it is not magic dust. It encodes a claim about the world — “these transformations leave the answer unchanged.” When that claim is true and complete, augmentation builds a model that reads the map from any angle. When it is partial, or when it asserts a symmetry the task does not have, it teaches the model to ignore something it should have learned. Rotations helped. Mirrors, applied carelessly, can hurt. The discipline is knowing which symmetries are real.