Boundary-Guided Trajectory Prediction for Road Aware and Physically Feasible Autonomous Driving

This paper proposes a novel boundary-guided trajectory prediction framework that leverages HD map constraints and kinematic acceleration profiles to generate physically feasible, on-road, and robust autonomous driving predictions, significantly reducing off-road errors and improving generalization compared to existing baselines.

The Gist

Imagine you are teaching a robot to drive a car. The robot has to guess what other cars, pedestrians, and cyclists will do next. This is called trajectory prediction.

If the robot guesses wrong, it might think a car is going to drive through a wall, or worse, it might tell the car to drive off a cliff. Current AI models are getting really good at guessing, but they still make two big mistakes:

1. They get lost: They sometimes predict a car will drive through a park or a building because they don't "see" the road boundaries clearly.
2. They break physics: They might predict a car will turn instantly at 90 degrees or stop in mid-air, which is physically impossible for a real car.

This paper introduces a new way to teach the robot how to drive that fixes both problems. Here is how it works, using some simple analogies.

1. The "Train Track" Analogy (Solving the "Getting Lost" Problem)

Imagine you are trying to draw a path for a toy car.

• Old Way: You just tell the AI, "Go to that point over there." The AI draws a line. Sometimes, the line goes straight through a tree because the AI didn't know the tree was there.
• This Paper's Way: Instead of just giving a destination, the AI is given two invisible rails: a left rail and a right rail. Think of these rails as the edges of a train track.
  • The AI is told: "You can drive anywhere between these two rails, but you cannot cross them."
  • The AI learns to draw a path that floats perfectly in the middle of these rails. Even if the road curves or splits, the AI knows exactly where the "safe zone" is because it's holding onto the rails.

The Magic Trick: The AI doesn't just pick one path. It learns to "mix" the left rail and the right rail. If the car needs to stay on the left side of the lane, the AI leans the path toward the left rail. If it needs to move right, it leans toward the right rail. This ensures the car never drives off-road.

2. The "Bicycle vs. Skateboard" Analogy (Solving the "Breaking Physics" Problem)

Imagine you are pushing a skateboard and a bicycle.

• The Skateboard (Old AI): It can spin 360 degrees instantly. It can stop in zero distance. It's great for drawing, but terrible for driving a real car.
• The Bicycle (This Paper's AI): A bicycle has rules. You can't turn the handlebars 90 degrees instantly, or you'll fall over. You have to lean and turn gradually.

This paper adds a special "Physics Filter" at the end of the AI's brain.

1. The AI first draws a rough path (the "Superposition Path") between the rails.
2. Then, it asks: "If I were a real car, could I actually drive this path?"
3. It calculates exactly how much to accelerate and how much to turn the steering wheel to make that path happen.
4. If the path requires a turn that is too sharp for a real car, the AI automatically smooths it out. It's like a "spell-check" for physics that fixes impossible moves before they happen.

3. The "Adversarial Attack" Test (The Stress Test)

The researchers wanted to see if their new method was tough. They used a technique called "Scene Attack," which is like a magic trick where they slightly distort the road map (making a straight road look like a wavy ripple) to confuse the AI.

• The Old AI (HPTR): When the road looked weird, the old AI panicked. It thought, "I don't know where the road is!" and 66% of the time, it predicted the car would drive off into the bushes.
• The New AI: Because it was holding onto those "invisible rails" (the boundaries), it didn't care if the road looked wavy. It knew, "As long as I stay between these rails, I'm safe." It kept the car on the road 99% of the time, even when the map was distorted.

Summary: Why is this a big deal?

Think of this new system as giving the self-driving car a guardrail and a physics textbook at the same time.

• Guardrails: It keeps the car on the road, even in confusing situations or when the map data is messy.
• Physics Textbook: It ensures the car moves like a real vehicle, not like a video game character that can teleport or spin instantly.

The result is a self-driving system that is safer, more reliable, and much less likely to make "silly" mistakes that could cause accidents. It's a step closer to having a robot driver that you can actually trust with your life.

Technical Summary

Here is a detailed technical summary of the paper "Boundary-Guided Trajectory Prediction for Road Aware and Physically Feasible Autonomous Driving."

1. Problem Statement

Accurate trajectory prediction is critical for autonomous driving safety. While deep learning models have improved performance, they face two persistent challenges:

• Lack of Road Awareness: Models often predict trajectories that go "off-road" (violating road topology) or fail to generalize to unseen road layouts. Existing methods like goal-conditioned approaches often only constrain the final point, ignoring intermediate path continuity. Set-based methods lack resolution and flexibility.
• Physical Infeasibility: Many models generate trajectories with unrealistic accelerations or curvatures because they lack explicit kinematic constraints during the regression process. This leads to predictions that a real vehicle cannot execute.

Current solutions often involve trade-offs between complexity, flexibility, and the guarantee of physical feasibility.

2. Methodology

The authors propose a novel framework that formulates trajectory prediction as a constrained regression task guided by permissible driving directions and their boundaries. The architecture consists of four main components:

A. Boundary Set Generation

Instead of relying on pre-defined trajectory sets or single-lane Frenet frames, the system dynamically extracts boundary sets from High-Definition (HD) maps based on the agent's current state.

• Algorithm: It identifies start lanes, performs reachability analysis to find goal lanes, and clusters them by driving direction.
• Extraction: Using a modified Depth-First Search (DFS), it extracts the leftmost and rightmost lanes for each permissible driving direction.
• Representation: These boundaries are represented as polylines (left and right edges). The network learns to predict a "superposition path" by weighting points between these left and right boundaries, ensuring the entire trajectory remains on-road.

B. Network Architecture

The model uses HPTR (Heterogeneous Polyline Transformer) as a backbone, enhanced with the boundary set:

• Input: Agent history, map lanes, and the generated boundary set (up to $N_b$ boundaries per scene).
• Encoder: Uses a Polyline Encoder (PointNet-based) to process map and agent data. A Transformer-based Interaction Encoder (using KNARPE attention) captures interactions between agents, maps, and boundary segments.
• Boundary Decoder: Combines boundary embeddings with agent embeddings. It uses learnable anchor tokens to generate multi-modal representations, preventing mode collapse and enabling diverse trajectory predictions.

C. Prediction Head & Superposition

• Superposition Weights: The network predicts weights for left and right boundary points for each mode. These weights sum to one, mathematically guaranteeing the generated path lies strictly between the road boundaries.
• Acceleration Profile: The model predicts a time-step acceleration profile constrained within physical limits ($-a_{max}$ to $a_{max}$).

D. Pure Pursuit Layer (Kinematic Constraint)

A parameter-free Pure Pursuit layer converts the superimposed path and acceleration profile into a final trajectory.

• It uses a differential kinematic model to update the vehicle state.
• It enforces a maximum curvature limit ($\kappa_{max}$) to ensure the trajectory is physically feasible for the specific vehicle dynamics.
• This layer acts as a final filter, eliminating any physically impossible turns or accelerations.

3. Key Contributions

1. Boundary Extraction Algorithm: A novel method to extract permissible driving direction boundaries from HD maps, overcoming the limitations of single-lane Frenet approaches and allowing for complex maneuvers like lane changes and U-turns.
2. Constrained Regression Framework: Reformulating trajectory prediction as a regression between left and right boundaries, ensuring 100% on-road adherence by design.
3. Physically Feasible Generation: Integrating a Pure Pursuit kinematic layer and acceleration constraints to guarantee that all predicted trajectories are executable by a real vehicle.
4. Robust Generalization: Demonstrating superior performance in unseen road topologies and under adversarial attacks compared to state-of-the-art baselines.

4. Experimental Results

The model was evaluated on the Argoverse-2 (AV2) dataset against the HPTR baseline and HPTR kin (HPTR with kinematic constraints).

• Accuracy: The proposed model (HPTR bd) shows a slight decrease in standard benchmark metrics (minADE/minFDE) compared to the unconstrained HPTR but significantly outperforms HPTR kin. Notably, it achieves a lower Final Displacement Error (minFDE) than the original HPTR, indicating better goal prediction.
• Physical Feasibility:
  • HPTR: 89.1% of predicted trajectories contained at least one infeasible step.
  • HPTR bd (Ours): 0.0% infeasible steps. The kinematic layer completely eliminated physically impossible predictions.
• Generalization & Robustness (Scene Attacks):
  • Under "Scene Attack" perturbations (simulating unseen road topologies), the baseline HPTR failed catastrophically, with 66% of trajectories going off-road.
  • The proposed model reduced the off-road rate to just 1%, demonstrating exceptional robustness to unseen road layouts.
• Complex Maneuvers: The model outperformed baselines in complex scenarios like U-turns, which are rare in training data, proving its ability to generalize without overfitting to simple straight-line movements.

5. Significance

This work addresses the critical safety gap in autonomous driving prediction systems. By shifting from unconstrained regression to boundary-guided constrained regression, the authors ensure that predictions are not only accurate but also legally compliant (on-road) and physically executable.

The approach offers a flexible alternative to rigid set-based methods, capable of handling complex, multi-modal driving behaviors while maintaining strict adherence to physical laws. The drastic reduction in off-road predictions under adversarial conditions highlights the model's potential for deployment in real-world, safety-critical autonomous systems where robustness to unseen environments is paramount.