I-24 MOTION Scenario Dataset (I24-MSD)

• I-24 MOTION Scenario Dataset is a large-scale, scenario-based vehicle trajectory benchmark designed to retain realistic sensor noise for microscopic traffic simulation.
• The dataset packages map-aligned, multi-agent scenarios (up to 9 seconds at 10Hz) to replicate practical roadway sensing with inherent imperfections.
• Its noise-aware curation and evaluation framework demonstrate that models using techniques like label smoothing and focal loss outperform traditional baselines.

The I-24 MOTION Scenario Dataset (I24-MSD) is a large-scale, curated, scenario-based vehicle trajectory dataset for microscopic traffic simulation derived from the I-24 MOTION testbed on Interstate 24 near Nashville, Tennessee. Its defining design choice is that it intentionally preserves realistic sensor imperfections rather than fully sanitizing them, so that generative agent-based models are trained and evaluated under conditions closer to practical roadway sensing. The dataset is structured to be compatible with autonomous-driving motion datasets, includes map-aligned context, and is accompanied by noise-aware modeling and evaluation procedures that use modern generative simulation metrics (Jayawardana et al., 10 Aug 2025).

1. Origin within the I-24 MOTION measurement ecosystem

I24-MSD is built on the broader I-24 MOTION infrastructure, which consists of 276 pole-mounted high-resolution traffic cameras providing seamless coverage of approximately 4.2 miles of Interstate 24, with trajectories extracted from video imagery using computer vision techniques (Gloudemans et al., 2023). That underlying instrument was designed for freeway traffic science and produces lane-resolved, road-aligned vehicle trajectories at scale; I24-MSD repackages this sensing capability into a scenario-based format tailored to microscopic simulation rather than direct trajectory archive consumption.

The motivation for this reformulation is specific to infrastructure-origin sensing. Unlike vehicle-mounted sensing, infrastructure-based sensing surfaces a messier combination of occlusions, calibration errors, temporal discontinuities, and map inconsistencies. I24-MSD treats these artifacts as part of the learning problem. In that sense, the dataset is not merely a cleaned subset of I-24 MOTION trajectories; it is a benchmark that operationalizes the claim that practical microsimulation should be robust to the kinds of imperfections transportation engineers actually face in deployed camera systems (Jayawardana et al., 10 Aug 2025).

A plausible implication is that I24-MSD occupies a methodological middle ground between raw infrastructure trajectories and heavily curated “ground-truth-like” motion datasets. This positioning is central to its relevance for simulation research: it is intended neither as a pure sensing benchmark nor as a fully idealized behavioral dataset.

2. Scenario representation, schema, and scale

Each sample in I24-MSD is a “traffic scenario” up to 9 seconds long, recorded at 10Hz, and each scenario includes up to 32 vehicle trajectories (user-configurable). For every agent, the dataset provides per-timestep $x, y, z$ coordinates and heading. Each trajectory is represented as a per-timestep sequence with 10 samples/sec $\times$ up to 9 sec = 90 samples per agent. The dataset also includes vectorized road maps aligned to each scenario, containing lanes and static infrastructure data, and uses the TFRecord storage format for compatibility with AV simulation datasets such as Waymo Open Motion (Jayawardana et al., 10 Aug 2025).

Aspect	Specification	Notes
Scenario unit	Up to 9 seconds at 10Hz	Long-horizon, multi-agent prediction
Agent state	$x, y, z$, heading	Up to 90 samples per agent
Context and format	Vectorized road maps; TFRecord	Lanes, static infrastructure, AV-dataset compatibility
Splits	Train/Validation noisy; Test lower-noise	Test approximates “ground truth” for evaluation

In the comparison reported with other motion datasets, I24-MSD is listed with 3.29m tracks, 6.8 seconds average length, 8 seconds horizon, 570k scenarios, 40 total hours, 6.5km of roadways, 10Hz sampling rate, and 1 object type (Jayawardana et al., 10 Aug 2025). The same comparison places it alongside Lyft, NuScenes, Argoverse, Interaction, and Waymo, but its design emphasis differs: it is scenario-based and map-aligned like AV datasets while preserving the sensing imperfections typical of freeway camera systems.

The split design is methodologically important. Train/Validation splits contain naturally noisy data representative of real-world sensing, whereas the Test set is more curated and lower-noise, serving as an approximation to evaluation-grade reference data. This means the dataset simultaneously supports robustness-oriented training and controlled benchmarking.

3. Preservation of sensor imperfection and curation strategy

The most distinctive property of I24-MSD is the intentional retention of noise. The documented sources of imperfection include camera pole movement (thermal, wind-induced), occlusions, motion blur, dust, glare, frame drops and skipped frames, multi-camera calibration and synchronization mismatches that induce ID switches, fragmented tracks, false positives/negatives, and map misalignments such as vehicles appearing off-lane, outdated lane information, or unannotated detours and road closures (Jayawardana et al., 10 Aug 2025).

These artifacts produce jittered, noisy, incomplete, or sometimes implausible trajectories, including examples such as overlapping vehicles, off-road movement, and abrupt heading changes. The dataset is therefore post-processed, but not fully scrubbed. Its curation strategy explicitly avoids over-sanitization.

The post-processing uses heuristic filtering to remove clearly unusable cases, including completely off-road vehicles, physically implausible behavior such as instant multi-lane shifts, tiny/fragmented tracks, and trajectory overlaps (i.e., collisions); it also densifies map polylines. At the same time, vehicles grazing edges or plausible edge cases are retained. The stated rationale is that corrections are chosen to mirror what typical transportation engineers can practically perform, thereby reflecting practically achievable data quality (Jayawardana et al., 10 Aug 2025).

A common misconception is to treat such noise as merely a defect in dataset preparation. I24-MSD adopts the opposite position: realistic imperfection is part of the benchmark definition. The controversy is therefore not whether the data are noisy—they are—but whether simulation models should be trained on sanitized abstractions or on the operational quality level of infrastructure sensing. The dataset is explicitly aligned with the latter view.

4. Generative modeling formulation and noise-aware learning

The associated modeling framework casts microscopic simulation as a conditional generative modeling problem. The environment is formalized as a Hidden Markov Model (HMM) with masked latent world states,

$$\mathcal{H} = \left( \mathcal{S}, \mathcal{O}, p(o_t \mid s_t), p(s_t \mid s_{t-1}) \right),$$

where $\mathcal{S}$ denotes latent world states and $\mathcal{O}$ observable quantities. For all $N$ vehicles at time $t$, observations are written as $o_t = [o_t^{(1)}, \dots, o_t^{(N)}]$, and observation noise is introduced as

$\tilde{o}_t = \psi(o_t, \epsilon_t).$

This formulation makes the corruption process explicit rather than relegating it to preprocessing (Jayawardana et al., 10 Aug 2025).

The paper adapts an autoregressive transformer (SMART) using several noise-robust losses commonly used in vision and language modeling: Cross-Entropy Loss (CE), Label Smoothing (CE+LS), Focal Loss, and Symmetric Cross Entropy (SCE). The reported hyperparameters are label smoothing $\times$0, focal loss $\times$1, and SCE with $\times$2, $\times$3, $\times$4 (Jayawardana et al., 10 Aug 2025).

The substantive point is not only that the dataset contains noise, but that the training objective is modified to account for it. This suggests a broader methodological shift: robustness can be pursued through loss design and model formulation rather than through increasingly aggressive data cleaning. In the context of infrastructure-based traffic simulation, that distinction is consequential because many sensing artifacts are persistent features of deployment environments, not one-off annotation mistakes.

5. Evaluation protocol and reported performance

Evaluation follows metrics commonly used in AV simulation benchmarks. The reported measures are Realism, Kinematic, Interactive, Map-based, and minADE. Realism is described as measuring how realistic the distribution of simulated trajectories is, while the kinematic, interactive, and map-based scores provide behavioral breakdowns; minADE is the standard Minimum Average Displacement Error (Jayawardana et al., 10 Aug 2025).

The benchmark compares classical baselines with SMART variants. The reported results are:

• IDM: Realism 0.7001, Kinematic 0.7592, Interactive 0.8192, Map-Based 0.5365, minADE 4.06.
• Constant Speed: Realism 0.6891, Kinematic 0.7581, Interactive 0.7904, Map-Based 0.5429, minADE 4.22.
• SMART (CE): Realism 0.7698, Kinematic 0.7353, Interactive 0.8253, Map-Based 0.7183, minADE 2.01.
• SMART (CE+LS): Realism 0.7922, Kinematic 0.7406, Interactive 0.8300, Map-Based 0.7731, minADE 1.34.
• SMART (Focal): Realism 0.7896, Kinematic 0.7386, Interactive 0.8300, Map-Based 0.7667, minADE 1.44.
• SMART (SCE): Realism 0.7837, Kinematic 0.7382, Interactive 0.8281, Map-Based 0.7526, minADE 1.59 (Jayawardana et al., 10 Aug 2025).

The reported takeaway is that all SMART variants outperform traditional baselines, and that noise-aware losses, especially label smoothing and focal loss, improve realism, map compliance, and displacement error. The core empirical argument is therefore not merely that transformers outperform IDM or constant-speed baselines, but that engaging with data imperfection explicitly is effective. This is the principal evidence supporting the dataset’s curation philosophy.

A plausible implication is that benchmark progress on I24-MSD may depend less on eliminating all residual noise and more on improving uncertainty-aware sequence modeling, tokenization, and context conditioning under imperfect observations.

6. Relation to adjacent I-24 datasets and research significance

I24-MSD is one component of a broader I-24 MOTION data ecosystem. The foundational I-24 MOTION release provides roadway-aligned vehicle trajectories with fields such as timestamps, $\times$5 and $\times$6 position, dimensions, direction, and class, sampled at 25 Hz over a 4.2 mile freeway segment and intended for freeway traffic science (Gloudemans et al., 2023). The Interstate-24 3D Dataset addresses a different layer of the stack: it is a benchmark for 3D multi-camera vehicle tracking, with 877,000 labeled 3D boxes across 3 scenes, 57 minutes across all views, and baseline tracking pipelines that remain challenged by high-speed motion and heavy occlusion (Gloudemans et al., 2023). The virtual trajectory dataset, in turn, derives smooth, continuous trajectories from the I-24 MOTION INCEPTION data by constructing a macroscopic speed field and integrating

$\times$7

with applications to speed variability, travel times across lanes, traffic waves, and energy-related analysis (Ji et al., 2023).

Against that background, I24-MSD is specifically the scenario-based benchmark for microscopic simulation. It differs from the 25 Hz trajectory archive by packaging data into map-aligned, short-horizon, multi-agent episodes; it differs from I24-3D by emphasizing simulation rather than detection and tracking; and it differs from virtual trajectories by preserving sensing imperfections rather than smoothing them away. These contrasts clarify its niche within the same measurement program.

The reported significance of I24-MSD is that it provides a direct bridge from infrastructure-origin traffic data to modern generative simulation pipelines. Because it adopts the same general scenario structure and storage conventions used in AV motion datasets, it enables reuse of existing transformer-based modeling tools and metrics while exposing them to the operational noise profile of freeway camera sensing (Jayawardana et al., 10 Aug 2025). The dataset is also described as open-source and extensible, with code for user-custom reprocessing and public availability at https://ct135.github.io/i24-msd/.

This suggests a broader role for I24-MSD in research on noise-aware model architectures, reinforcement learning with noisy data, and practical microsimulation benchmarks. More narrowly, it establishes a test case for whether realistic traffic simulation can be learned from infrastructure data without first converting that data into an idealized, nearly noise-free surrogate.