💻 computer science

PACE: Polar Axis-Conditioned Estimation for PairUAV Relative Localization

The paper introduces PACE, a method for PairUAV relative localization that improves performance by assigning axis-specific readout interfaces to better capture the distinct structural preferences of heading and range estimation, ultimately achieving a final deterministic challenge score of 0.001874.

Original authors: Ze Rong

Published 2026-07-21

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Ze Rong

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 you are standing on a rooftop with a drone, looking at another drone flying nearby. You need to tell your drone exactly where to go to meet the other one. You don't need a complex map; you just need two simple instructions: "Turn this much to the left or right" (the heading) and "Fly this far forward" (the range). This is the heart of a field called relative localization. It's like teaching a robot to play a game of "Red Rover" with another robot, using only their cameras to figure out where the other one is.

For a long time, scientists have tried to build AI models that can look at two pictures and instantly spit out these two numbers. The usual approach is to treat the "turn" and the "distance" as a single package, forcing the AI to use the exact same brain cells to figure out both. But what if the brain works better if it uses different parts for different tasks? What if the part of the brain good at spotting angles is actually bad at judging distance, and vice versa? This paper dives into that question, asking whether we should stop treating these two directions as identical twins and start treating them like distinct siblings with different strengths.

The Problem: One Size Does Not Fit All

The researchers behind this study, working on a challenge called PairUAV, noticed something strange. They were training an AI to look at two drone photos and predict a direction and a distance. They assumed that since both numbers describe the same relationship between the drones, the AI should use the same "decoder" (the final layer of its brain) to figure out both.

But when they peeked under the hood, they found a mismatch. It turned out that the AI's "best guess" for the angle was often coming from a different stage of its learning process than its "best guess" for the distance. In fact, on a test run, the AI's favorite settings for getting the angle right disagreed with its favorite settings for getting the distance right 80.8% of the time!

It's like trying to bake a cake and a loaf of bread at the same time using the exact same oven temperature and timer. You might get a decent result for one, but the other will likely be a disaster. The paper argues that forcing the AI to use the same "evidence" (the internal clues it gathers from the images) for both tasks is holding it back.

The Solution: PACE (Polar Axis-Conditioned Estimation)

To fix this, the authors introduced a new method called PACE. Think of PACE as a smart kitchen manager who realizes that while the chef (the main AI) is the same, the tools used for the cake and the bread should be different.

Instead of one single output line, PACE splits the final step into two specialized paths:

The Angle Path: This path is allowed to look at a mix of "middle" and "late" clues from the AI's brain. It's like letting the angle-finder peek at both the rough sketch and the final details of the drawing.
The Distance Path: This path is kept simple and direct, sticking only to the "late" clues (the final, most polished details). It's like telling the distance-finder to ignore the rough drafts and only trust the finished blueprint.

The paper tested two versions of this idea:

MDHR: A version that lets the angle path grab extra clues from the middle of the process.
PAAER: A version that protects the distance path, keeping it strictly on the direct line to the final clues, while the angle path gets a bit more freedom.

What They Found

The results were clear: treating the two directions differently worked better than treating them the same.

The Mismatch is Real: The study confirmed that the "best" version of the AI for angles is rarely the "best" version for distance. They are constantly out of sync.
Better Scores: By using this split approach, the AI got much better at its job. The best single AI model they released (MDHR) scored 0.002460 on the official test. Another version (PAAER) scored 0.002514, but it was slightly better at getting the angle right.
The "Tail" Problem: They also noticed that the distance predictions sometimes had a "high-error tail"—meaning most of the time it was great, but occasionally it would make a huge mistake. By adding a special "tail-weighted" training step, they managed to reduce those massive errors by over 95%.

The Final Trick: Deterministic Packaging

The paper also mentions a final step called "deterministic packaging." This isn't a new learning trick; it's more like a final polish. After the AI makes its guess, the system takes the predictions from three different models, mixes them together with specific math (like a secret recipe: 51.1% of one, 18.9% of another, and 30% of a third), and then snaps the numbers to the nearest allowed values (rounding the angle to the nearest 2 degrees and the distance to the nearest available number).

When they combined the smart split-brain AI (PACE) with this final mixing and rounding trick, they achieved a final score of 0.001874. This is the number that won the challenge, but the authors are careful to say this score comes from the whole package of smart learning plus careful math, not just the learning part alone.

The Takeaway

The main lesson here is that even when two things seem like they belong together (like turning and flying forward), they might need different kinds of help to get it right. The paper doesn't claim to have solved everything or found a magic bullet that works for every robot in the universe. Instead, it suggests that in the specific world of drone-to-drone navigation, letting the AI use different "evidence" for different directions is a much smarter way to go.

The authors are honest about the limits: their "mixing recipe" was tuned specifically for this one competition and might not work elsewhere. But the core idea—that we should stop forcing AI to use the same brain cells for everything—seems to be a solid step forward for making drones smarter and more precise.

Technical Summary: PACE – Polar Axis-Conditioned Estimation for PairUAV Relative Localization

Problem Statement

The PairUAV relative localization task requires a model to infer a polar navigation command—a heading angle ( $\theta$ ) and a metric range ( $d$ )—from a pair of UAV-view images. While both outputs describe the same pairwise pose, standard end-to-end joint regression treats them as homogeneous coordinates. This approach forces both outputs to consume the same decoder evidence and rely on the same optimization state. The authors argue that this assumption is flawed: controlled readout probes reveal that heading and range favor different decoder-depth combinations, their optimal training checkpoints often disagree, and range errors exhibit a distinct high-error tail that differs from heading errors.

Methodology

The paper introduces PACE (Polar Axis-Conditioned Estimation), a framework that retains a shared image-pair representation (based on the Reloc3r-style encoder-decoder) but decouples the readout interfaces for the two polar axes.

1. Axis-Conditioned Readout:
Instead of a single terminal interface, PACE assigns specific evidence subsets to each axis:

Heading ( $\theta$ ): Aggregates relational evidence from middle and late decoder layers.
Range ( $d$ ): Remains attached to a direct late metric interface (using only the final layer features).
Shared Backbone: The encoder-decoder backbone remains shared; the method does not claim latent factor independence but rather optimizes evidence allocation.

2. Specific Implementations:

Multi-Depth Heading Readout (MDHR): Uses the terminal range path and aggregates selected intermediate and final decoder features specifically for the heading head.
Protected Axis-Asymmetric Expert Readout (PAAER): A complementary instantiation where range is predicted directly from the late layer ( $z(L)$ ), while heading utilizes a query-bridge attention mechanism over middle and late layers. The term "expert" here refers to an axis-specialized readout, not a mixture-of-experts routing mechanism.

3. Optimization Strategy:
The models are fine-tuned using a metric-aware objective function. The loss functions for heading and range are normalized by the target magnitude to handle scale variations.

Range Loss: Includes a SmoothL1 term. In the final PAAER continuation, a sample weight is applied to range terms based on distance magnitude (increasing weight for distances $>80$ units) to specifically address high-error tails.
Training Protocol: Models start from Reloc3r-512 checkpoints, use 512×384 inputs, and undergo full-model fine-tuning.

4. Deterministic Challenge Packaging:
The final submission is not a single learned model but a deterministic fusion of three predictors:

Fusion: A weighted sum of the PAAER range, an MDHR range, and a metric-aware Reloc3r baseline range.
Calibration: The heading is rounded to the nearest $2^\circ$ lattice, and the fused range is projected to the nearest value in the public training/dev support set.
Note: These operations are explicitly described as "packaging" rather than learned contributions.

Key Contributions

Identification of Axis-Conditioned Reliability: Through matched readout controls and cross-surface trajectory diagnostics, the authors demonstrate that heading and range exhibit different reliability profiles regarding decoder depth and training checkpoints (disagreeing on 80.8% of a validation trajectory).
PACE Framework: Proposes a method that preserves shared pair context while allocating decoder evidence specifically by polar axis, utilizing MDHR and PAAER as instantiations.
Separate Evaluation: Evaluates learned estimators and the deterministic submission protocol separately, releasing artifacts to reconstruct the reported challenge result.

Results

The paper reports results on the official hidden test set of the PairUAV challenge:

Raw Predictors: The strongest released raw predictor is MDHR, achieving a final score of 0.002460. PAAER scores 0.002514, showing a slightly lower angle relative error (0.002636 vs. 0.002646 for MDHR) but a higher range error.
Final Submission: The deterministic fusion and calibration package (PACE final) achieves a score of 0.001874, significantly outperforming the raw learned predictors and the official baseline (0.633957).
Error Analysis: Range errors in standard models exhibit a severe high-error tail. A tail-weighted continuation of PAAER reduced the P95, P99, and maximum range errors by 85.3%, 97.1%, and 95.2% respectively, though this was a specific diagnostic continuation rather than a general ablation.

Significance and Claims

The paper claims that PairUAV relative localization benefits from a shared image-pair representation, but its polar axes do not need to inherit the same decoder evidence or training state. The significance lies in demonstrating that treating evidence utility and optimization reliability as homogeneous is a failure point.

The authors maintain a modest scope:

They do not claim hard latent factor disentanglement or complete gradient isolation.
The performance gains from the final submission are attributed to "challenge engineering" (deterministic fusion and lattice snapping) rather than a transferable learned rule.
The observed axis-dependent behaviors (tails, checkpoint disagreement) are noted as specific to the PairUAV task structure and evaluator metrics, cautioning against generalization to other benchmarks without controlled follow-ups.
The work serves as a bounded conclusion that in shared pairwise representations, each predicted coordinate should read the evidence on which it is most reliable.