💻 computer science

Safer Trajectory Planning with CBF-guided Diffusion Model for Unmanned Aerial Vehicles

This paper presents AeroTrajGen, a novel diffusion-based framework for unmanned aerial vehicles that integrates control barrier function (CBF)-guided sampling during inference to generate safe, agile, and diverse aerobatic trajectories while significantly reducing collision rates and reliance on safety-verified training data.

Original authors: Peiwen Yang, Shiyu Bai, Weisong Wen, Yixin Gao, Jiahao Hu

Published 2026-04-21

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Peiwen Yang, Shiyu Bai, Weisong Wen, Yixin Gao, Jiahao Hu

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 teaching a robot drone to perform amazing, high-speed aerobatic stunts—like loops, rolls, and spirals—inside a room filled with floating balloons (obstacles).

The challenge? You want the drone to be agile (doing cool tricks) and safe (not popping the balloons).

This paper introduces a new system called AeroTrajGen that solves this problem by combining two powerful ideas: a "creative artist" AI and a "strict safety guard."

Here is how it works, broken down into simple concepts:

1. The Creative Artist: The Diffusion Model

Think of a Diffusion Model like an artist who learns to draw by starting with a messy scribble and slowly cleaning it up until it becomes a perfect picture.

How it works: The AI has watched 2,000 videos of expert pilots flying drones. It learned the "feel" of how a drone should move to do a loop or a roll.
The Problem: When you ask this AI to draw a new path, it's very creative and diverse, but it's also a bit reckless. It might draw a path that goes straight through a wall or a balloon because it's focused on making the shape look cool, not on whether the path is physically possible or safe. It's like a talented painter who doesn't know how to avoid the furniture in the room.

2. The Strict Safety Guard: Control Barrier Functions (CBF)

Now, imagine a Safety Guard standing next to the artist. This guard knows the exact rules of the room: "Do not go within 1 meter of any balloon."

The Old Way: Usually, you would let the artist draw the picture first, and then the guard would look at it and say, "No, that hits a balloon! Start over!" This is slow and frustrating.
The New Way (AeroTrajGen): This paper puts the guard inside the artist's brain. As the artist is drawing the path, the guard is whispering in their ear: "Hey, you're getting too close to that balloon! Pull your hand slightly to the left."
The Result: The artist never even draws the dangerous part. The path is generated safely from the very first stroke.

3. The Magic Combination: "CBF-Guided Sampling"

The paper calls this CBF-guided sampling.

The Metaphor: Imagine you are walking through a dark forest (the AI generating a path). You have a flashlight (the AI's creativity) that shows you many possible paths. But you also have a magnetic compass (the Safety Guard) that pulls you away from the cliffs (obstacles).
How it helps: Without the compass, you might wander off a cliff because the path looked interesting. With the compass, you are gently steered toward safe ground while you are walking, so you never fall off.

What Did They Achieve?

The researchers tested this system in a computer simulation with 14 different types of crazy stunts (like "barrel rolls" and "figure-eights") and rooms full of obstacles.

The Result: The "unguided" AI (the artist without the guard) crashed into obstacles 99% of the time.
The Fix: The "CBF-guided" AI (artist + guard) crashed 0% of the time.
The Trade-off: The safe paths were slightly less "perfect" in terms of hitting the exact target spot compared to the reckless paths, but they were alive and safe. In the real world, being safe is much more important than being perfect.

Why Does This Matter?

Currently, if you want a drone to fly safely, you usually need to feed it thousands of examples of perfectly safe flights. That's hard to get.
This new method is special because:

It can learn from "messy" or risky data.
It adds the safety rules at the moment of creation (during inference), so it doesn't need to be retrained every time the room changes.
It allows drones to do complex, fun, and fast maneuvers without crashing, making them ready for real-world jobs like search and rescue or delivery in crowded cities.

In short: They taught a creative AI how to be a responsible pilot by giving it a built-in safety instinct, ensuring it can perform amazing stunts without hitting anything.

1. Problem Statement

Autonomous Unmanned Aerial Vehicles (UAVs) require trajectory planning that is both agile (capable of complex aerobatic maneuvers) and safe (collision-free). While Generative AI, specifically Diffusion Models, has shown promise in generating diverse and high-quality trajectories, they suffer from two critical limitations in safety-critical domains:

Unpredictability: Diffusion models can generate outputs that violate safety constraints (e.g., colliding with obstacles) because they are trained on data distributions without explicit hard safety guarantees.
Data Dependency: Ensuring safety often requires training on extensively safety-verified datasets, which are difficult and expensive to acquire.

The paper addresses the challenge of generating complex aerobatic trajectories in obstacle-laden environments without relying on post-hoc safety filters or massive amounts of pre-verified safe training data.

2. Methodology

The authors propose AeroTrajGen, a novel framework that integrates Control Barrier Functions (CBFs) directly into the inference phase of a diffusion model.

A. Core Architecture: Obstacle-Aware Diffusion Transformer

Model Type: A Transformer-based diffusion model designed for trajectory generation.
Input Conditioning: The model uses multi-modal conditioning to guide generation:
- Trajectory History: Past states to ensure continuity.
- Goal/Target: The desired endpoint.
- Maneuver Style: One-hot encoded labels for 14 distinct aerobatic maneuvers (e.g., barrel roll, Immelmann, figure-eight).
- Obstacle Context: An attention-based encoder processes static obstacle information (position and radius).
Training: Trained on a dataset of 2,000 expert demonstrations (augmented to 3,000) covering 14 maneuver styles. The loss function balances position accuracy, velocity, continuity with history, and acceleration smoothness.

B. CBF-Guided Sampling (The Innovation)

Instead of training the model to be safe or filtering outputs after generation, the authors modify the reverse diffusion process (sampling) to enforce safety constraints dynamically.

Integration: The standard score function of the diffusion model ( $\nabla_x \log p_t(x_t)$ ) is augmented with a safety gradient derived from CBFs ( $\nabla_x \log p_s(C | x_t)$ ).
Safety Probability Function: The authors define a safety probability $p_s$ $p_{s}$ based on the distance between the UAV and obstacles. They compare three formulations:
1. Exponential Barrier (Difference of Squares).
2. Normal CDF.
3. Logistic (Sigmoid) Function: Chosen as the optimal balance between smoothness and computational efficiency.
Guided Noise Prediction: The predicted noise ( $\epsilon$ ) is adjusted by adding a term proportional to the gradient of the barrier function. This "pushes" the generated trajectory away from obstacles during the denoising steps.
Time-Dependent Guidance: A guidance strength parameter $\gamma(t)$ is applied, which is strong in early diffusion steps (to steer the trajectory globally away from obstacles) and weaker in later steps (to allow fine-tuning of trajectory details).

3. Key Contributions

CBF-Guided Inference: A novel method to embed CBF safety constraints directly into the gradient flow of the reverse-time Stochastic Differential Equation (SDE) of a diffusion model, eliminating the need for retraining with safety-verified data.
Obstacle-Aware Transformer: A specialized architecture that jointly encodes environmental context (obstacles) and maneuver styles, enabling the generation of physics-informed, multi-style trajectories.
Safety-Accuracy Trade-off Analysis: A rigorous evaluation of different safety probability functions and guidance strengths, demonstrating how to tune the system for either maximum safety (zero collisions) or high trajectory fidelity.
Open-Source Framework: The release of the code and the AeroTrajGen framework for the research community.

4. Experimental Results

The model was evaluated in simulation with 3–5 randomly placed spherical obstacles.

Collision Reduction:
- Unguided Baseline: Suffered from a 99.0% collision rate in obstacle-rich environments (without the obstacle encoder) or 12.0% (with encoder but no CBF).
- CBF-Guided (Config C): Achieved a 0.0% collision rate using the Logistic safety function with strong guidance ( $\gamma = 4000$ ). This represents a 94.7% to 100% reduction in collision rates compared to baselines.
Safety vs. Accuracy Trade-off:
- While CBF guidance ensured zero collisions, it introduced a trade-off: Position RMSE increased from 0.47m (unguided) to 1.19m (CBF-guided).
- Maneuver Success: The overall success rate (collision-free + accuracy + target reach) for the best safety configuration was 42.0%, compared to 47.5% for a moderate guidance configuration. Complex maneuvers (e.g., Sphinx, Figure-Eight) saw lower success rates due to the conflict between high precision requirements and strict safety constraints.
Ablation Study:
- The Obstacle Encoder was found critical; without it, the unguided model had a 99% collision rate.
- The Logistic Safety Function with high guidance strength provided the most robust safety guarantees.

5. Significance

This paper bridges the gap between expressive generative models and safety-critical robotics.

Paradigm Shift: It moves safety enforcement from the training phase (requiring massive safe datasets) or post-processing (which can degrade trajectory quality) to the inference phase via mathematical guidance.
Practical Application: It demonstrates that diffusion models can be safely deployed for complex, high-speed aerobatic maneuvers in dynamic environments, provided safety constraints are integrated into the generative process.
Future Direction: The work lays the groundwork for adaptive constraint weighting and handling dynamic obstacles, offering a unified solution for safe autonomous flight.

In summary, AeroTrajGen proves that by guiding diffusion sampling with Control Barrier Functions, UAVs can generate complex, agile trajectories that are mathematically guaranteed to be collision-free, significantly advancing the state of safe autonomous aerial robotics.