Target-Bench: Can Video World Models Achieve Mapless Path Planning with Semantic Targets?

1 Introduction

• Existing evaluations emphasize visual fidelity and spatio-temporal consistency, leaving mapless planning that requires semantic understanding difficult to assess.The paper frames quantitative evaluation of video world models' reasoning, task-solving, and planning as an open need.Source: §1 · p. 2
• Target-Bench evaluates video world models on mapless path planning toward textual semantic targets in unstructured real-world environments.Its modular, replaceable world-decoder and metrics target semantic reasoning, spatial estimation, and planning.Source: §1 · pp. 2–3
• The dataset contains 450 quadruped-robot scenarios across 47 categories, with trajectories and human annotations for explicit and implicit targets.It comprises 112,500 frames from diverse indoor and outdoor environments.Source: §1 · p. 3
• Fine-tuning an open-source model on a small real-world robot dataset substantially boosts task-level planning performance.The paper presents this as a comparative study of open-source and proprietary models and reports data-efficient adaptation.Source: §1 · p. 3

World Models

• Foundation-scale video world models have improved prompt adherence and dynamics, but real-world robot task planning remains largely untested.Earlier work includes latent state-space models and newer video and interactive world models.Source: §2 · p. 3

Semantic Navigation Datasets

• Existing navigation datasets often lack explicit navigation targets, and some trajectory annotations are only up-to-scale.The paper contrasts Target-Bench's target-oriented real-robot teleoperation data with prior datasets.Source: §2 · p. 3, Table 1 · p. 4

Benchmarks for World Models

• Several world-model evaluation frameworks are confined to simulation, while robotics-oriented suites rely on subjective VLM judges that limit reproducibility.The paper also notes that existing benchmarks tend to emphasize explicit rather than implicit semantic instructions.Source: §2 · pp. 3–4

3 Target-Bench

• Target-Bench evaluates video-model semantic navigation in real-world robot scenes, including implicit targets and invariant task-driven scoring without subjective model judgments.Its pipeline reconstructs paths from generated video and scores target proximity and directional consistency.Source: §3 · p. 4, Table 2 · p. 8

3.2 Target Benchmark Architecture

• The benchmark pairs a world-decoder that extracts camera poses from generated videos with path evaluation that compares the resulting path against ground truth.The evaluation focuses on approach tendency rather than requiring an exact trajectory match.Source: §3 · p. 4, §3.2 · pp. 6–7

3.1 Target Dataset Setup

• The dataset contains 450 collected samples; the robot dog starts with the target in view, and a human operator uses direct goal-approach motion at constant forward speed, minimizing heading oscillations.Of these samples, 125 are curated as the benchmark set, while the remaining 325 are used exclusively for training.Source: §3.1.2 · p. 6

3.2.1 World-Decoder

• VGGT, SpaTracker, and ViPE reconstruct camera poses; monocular methods use displacement-based scale recovery, while ViPE outputs metric-scale poses through visual-inertial fusion.A single recovered scale factor rescales predicted translations while preserving relative geometry.Source: §3.2.1 · pp. 7–8

3.2.2 Path Evaluation

• Five path metrics assess accuracy, goal-reaching, and consistency, while the weighted score emphasizes endpoint proximity and approach consistency over exact path overlap.The ground-truth path is a rough motion-tendency reference, allowing different trajectories with similar tendencies to score well.Source: §3.2.2 · pp. 8, 10, Figure 6 · p. 9

4 Experiments

• The evaluation covers Sora 2, Veo 3.1 variants, Wan-series models, and fine-tuned Wan2.2-TI2V-5B models.Source: §4 · pp. 10–11

4.2 Evaluation for off-the-shelf Models

• 0.341 is the best off-the-shelf weighted overall score; Wan2.2-Flash also records FDE 1.362m, ADE 1.005m, MR 38.75%, and SE 0.292.Performance on implicit targets closely matches that on explicit targets, with small model-specific fluctuations.Source: Table 3 · p. 11, §4.2 · pp. 11–12, Table 4 · p. 12

4.1 Implementation Details

• The experiments use LoRA fine-tuning, API access for closed-source models, and generated videos at 720p or 1080p with 5–10-second durations.Source: §4.1 · p. 11

4.3 Fine-tuned Models

• 0.330 versus 0.084 is the fine-tuned Wan2.2-5B score; with data augmentation, it exceeds the base model by more than 469% and achieves the best overall score.Augmentation creates four clips per scenario from shifted 121-frame windows while reusing captions.Source: Table 3 · p. 11, §4.3 · p. 12

4.4 Ablation Study

• 0.862 is VGGT’s weighted overall score on ground-truth videos; SpaTracker scores lower and ViPE weakest, supporting the framework’s replaceable reconstruction-tool design.Source: Figure 8 · p. 13, §4.4 · p. 13

4.5 Sensitivity to the Planning Horizon Length

• 8% higher weighted overall score follows a 6-to-4-second horizon reduction for Wan2.2-I2V-Flash; Wan2.2-I2V-Plus improves by more than 17% from 8 to 6 seconds.Weighted scores consistently improve as the planning horizon decreases.Source: §4.5 · pp. 13–14, Table 6 · p. 14

• 0.341 is the best off-the-shelf score, despite most models showing plausible motion tendencies and apparent target understanding in generated videos.Implicit semantic targets make the benchmark challenging for video world models.Source: §5 · p. 14
• Target-Bench jointly evaluates spatio-temporal consistency, semantic reasoning, and geometric path accuracy, unlike benchmarks that isolate visual quality or physics consistency.Blurred frames, temporal discontinuities, or spatial warping can reduce reconstructed 3D path accuracy.Source: §5 · p. 14
• The decoder extracts an executable path from predicted future frames, with closed-loop navigation in unstructured mapless environments described as a potential use.Source: §5 · p. 14

• Fine-tuning an open-source 5B-parameter model on 325 scenarios surpasses the best off-the-shelf video world models.The authors identify closed-loop robot replanning and latent memory for receding-horizon navigation as future work.Source: §6 · p. 15