🤖 AI

LAMP: Lane-Aligned Motion Primitives for Feasible Trajectory Prediction

This paper introduces LAMP, a topology-aware motion forecasting framework that utilizes VQ-VAE-learned motion primitives and a feasibility-aware selector to generate multimodal trajectories that strictly adhere to lane topology while maintaining behavioral diversity and prediction accuracy.

Original authors: Sangjin Han, Hoseong Jung, Jeongtae Her, Changhyun Choi, H. Jin Kim

Published 2026-06-26

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Sangjin Han, Hoseong Jung, Jeongtae Her, Changhyun Choi, H. Jin Kim

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 to drive a car. To drive safely, the robot needs to guess what other cars, pedestrians, and cyclists will do next. This is called "motion forecasting."

The problem with many current robot drivers is that while they are good at guessing where a car might end up, they are bad at guessing how it will get there. They might predict a car will suddenly drive through a sidewalk, jump over a fence, or drive the wrong way down a one-way street. These predictions are "physically impossible" or "illogical," making them useless for a safety system that needs to avoid crashes.

The paper introduces a new system called LAMP (Lane-Aligned Motion Primitives) to fix this. Here is how it works, using some everyday analogies:

1. The Old Way: Guessing the Destination

Most current systems try to predict the future by picking a few random "target points" (like a destination on a GPS).

The Flaw: It's like telling a taxi driver, "Go to the airport," but not caring about the route. The driver might take a path that goes through a park or a river because the system didn't check the road map. The robot gets the end point right but the path wrong.

2. The LAMP Solution: A Library of "Driving Recipes"

Instead of guessing a destination, LAMP uses a library of pre-learned "motion primitives." Think of these as driving recipes or dance moves.

The Library (VQ-VAE): The system first learns from millions of real driving videos. It doesn't just memorize start and end points; it memorizes the shape of the movement. It learns specific patterns like "a smooth left turn at medium speed," "a quick lane change," or "a slow stop."
The Analogy: Imagine a chef who has a library of perfect recipes for "spaghetti," "steak," and "salad." Instead of trying to invent a new dish from scratch every time, the chef picks the best recipe that fits the ingredients available. LAMP picks the best "driving recipe" that fits the current traffic situation.

3. The Safety Filter: The "Map Police"

Even with a library of good recipes, you might accidentally pick a recipe that doesn't fit the current kitchen. For example, trying to make a "boat" recipe when you are on land.

The Problem: Sometimes the system might pick a "driving recipe" that looks cool but is impossible on the current road (e.g., a sharp turn that would drive the car off the road).
The Solution (Feasibility Selector): Before the robot actually plans the drive, LAMP has a "Map Police" officer. This officer looks at the current road layout (the lanes) and checks the list of potential driving recipes.
The Action: If a recipe says "turn left into a wall" or "drive off the road," the Map Police throws it out immediately. They only keep the recipes that are physically possible and legally allowed on that specific stretch of road.

4. The Result: Safe and Diverse Options

By combining the library of "driving recipes" with the "Map Police" filter, LAMP produces a set of predictions that are:

Diverse: It doesn't just guess one path; it offers many different plausible ways the car could move (like a fast lane change vs. a slow merge).
Safe: Every single option offered is guaranteed to stay on the road and follow traffic rules.

Why This Matters

The paper tested this on a massive dataset of real driving scenarios (Argoverse 2). They found that LAMP is just as good at guessing the final location as the best existing systems, but it is much better at ensuring the path is realistic.

In short: LAMP stops the robot from daydreaming about driving through buildings. Instead, it gives the robot a menu of realistic, road-legal driving options, ensuring that whatever the robot decides to do next, it won't crash into a curb.

Technical Summary: LAMP (Lane-Aligned Motion Primitives) for Feasible Trajectory Prediction

1. Problem Statement

Motion forecasting is a critical component of autonomous driving systems, tasked with predicting the stochastic future trajectories of surrounding agents based on historical states and map context. While existing predictors effectively minimize standard displacement errors (e.g., ADE, FDE), they often fail to ensure that multimodal predictions adhere to lane topology.

Current approaches typically rely on predefined intention priors, such as trajectory anchors or target points, which are often designed independently of specific scene contexts. Consequently, these scene-agnostic priors frequently generate structurally infeasible trajectories, particularly for lower-probability modes. These violations of physical and logical constraints (e.g., off-road predictions or impossible lane transitions) render the prediction set unreliable for safety-critical downstream planning modules, which must evaluate risks and ensure safe decision-making.

2. Methodology

The authors propose LAMP (Lane-Aligned Motion Primitives), a topology-aware forecasting framework designed to anchor multimodal predictions to structured motion primitives aligned with lane topology. The framework is built upon the Motion Transformer (MTR) architecture within the UniTraj platform and consists of three core components:

A. Discrete Intention Queries via VQ-VAE

Instead of modeling driving intentions as static endpoints or continuous latent variables, LAMP formulates intention as a discrete set of learnable queries using a Vector Quantized Variational AutoEncoder (VQ-VAE).

Learning Process: A Non-Saturating Vector Quantization (NSVQ) based VQ-VAE is trained on collected motion trajectories. An encoder maps ground-truth trajectories into a latent space, which is quantized to the nearest code vector in a learnable codebook. A decoder reconstructs these codes into trajectory prototypes, referred to as motion primitives.
Query Formation: Each code in the discrete codebook is decoded into a motion primitive. An intention query is then formed by concatenating the code embedding with a spatial grounding term derived from the primitive's endpoint. This formulation captures rich spatiotemporal dynamics and recurring driving patterns (e.g., varying speeds on straight paths or different curvatures for turns) that endpoint-based methods often miss.

B. Feasibility-Aware Intention Selector

To ensure the generated hypotheses are physically and logically feasible, LAMP introduces a lane-topology-guided intention selector that operates prior to the final decoding stage.

Mechanism: The selector evaluates the $K$ motion primitives decoded from the intention queries. It utilizes a lane-topology prior, constructed by traversing HD-map lane connectivity from the agent's current pose to identify reachable lanes.
Scoring: A feasibility scoring function computes the distance between each hypothesis and the set of reachable lanes, generating a soft teacher distribution. The selector is trained to match this distribution, effectively filtering out unreachable or off-road intention queries.
Selection: Only the top- $L$ feasible intentions are retained and forwarded to the Transformer decoder. This prevents infeasible hypotheses from contaminating the attention context, ensuring the final prediction set consists of diverse yet topology-consistent trajectories.

C. Training Pipeline

The framework employs a two-stage training strategy:

Stage 1: The NSVQ-based VQ-VAE is trained to learn the discrete codebook of intention queries and motion primitives. The codebook and decoder are then frozen.
Stage 2: The forecasting backbone (MTR) and the intention selector are trained end-to-end. The total loss combines the standard negative log-likelihood loss (for the best-matching mode) with an auxiliary loss that minimizes the Kullback-Leibler divergence between the selector's predictions and the lane-topology prior.

3. Key Contributions

The paper outlines three primary contributions:

LAMP Framework: A topology-aware motion forecasting framework that improves the reliability of multimodal prediction sets by combining structured motion primitives with feasibility-aware intention selection.
Discrete Intention Formulation: A novel formulation of driving intention as discrete queries via a VQ-VAE, where each query decodes to a trajectory prototype that captures rich spatiotemporal dynamics beyond simple endpoints.
Lane-Topology-Guided Selection: The introduction of an intention selector that filters infeasible hypotheses based on lane topology, thereby improving feasibility and diversity metrics while preserving displacement accuracy.

4. Experimental Results

The authors evaluated LAMP on the Argoverse 2 Motion Forecasting Dataset, comparing it against strong baselines including MTR, Wayformer, Forecast-MAE, EMP, and Autobot.

Accuracy: LAMP achieves prediction accuracy comparable to state-of-the-art baselines on standard displacement metrics (b-minADE6, b-minFDE6).
Feasibility: The model demonstrates substantial improvements in structural reliability. Specifically, it outperforms baselines in DAC (Drivable Area Consistency) and FR (Feasibility Rate, which includes traffic rule constraints). The improvement in FR is particularly notable, indicating a reduction in traffic-rule-violating motions.
Diversity: LAMP improves diversity metrics (APD6, FPD6) while maintaining feasibility. The authors introduce DwF (Diversity-while-Feasible), a metric that measures diversity only over the feasible subset, where LAMP also shows significant gains.
Qualitative Analysis: Visualizations on intersection scenarios show that while baselines often concentrate on a single dominant mode or assign confidence to off-road trajectories, LAMP generates diverse, plausible maneuvers that strictly adhere to lane topology.

5. Significance and Claims

The paper claims that LAMP addresses a critical gap in autonomous driving perception: the trade-off between prediction accuracy and structural feasibility. By explicitly modeling driving intentions as learnable motion primitives and filtering them through a lane-topology prior, LAMP provides a reliable multimodal prediction set.

The significance lies in the framework's ability to generate hypotheses that are not only diverse (covering various driving intentions) but also planner-relevant (adhering to physical and logical constraints). This ensures that downstream planning modules receive a set of trajectories that can be safely evaluated for risk, reducing the likelihood of planning failures caused by infeasible predictions. The authors note that while map-adaptive decoding via LoRA was explored to further increase feasibility, it resulted in a collapse of multimodal diversity; thus, the final model prioritizes the balance between feasibility and diversity without early map conditioning.