NaviTrace: First-Person Navigation Traces for Embodied AI Evaluation

NaviTrace is a Visual Question Answering benchmark designed to test how well vision-language models understand real-world navigation. It pairs first-person outdoor images with natural language instructions, requiring models to predict a 2D sequence of waypoints an agent would follow. The dataset was created by Voxel51 and was last updated on May 26, 2026.

Use Cases

• Benchmarking model performance on embodied navigation tasks based on first-person images and instructions.
• Training models to predict 2D navigation traces based on scene understanding.
• Evaluating how different embodiment types affect navigation path planning.
• Developing agents that can interpret and follow natural language navigation commands in outdoor environments.

Strengths

• Specifically designed for evaluating embodiment-specific navigation understanding in vision-language models.
• Focuses on real-world, first-person outdoor scenes paired with natural language instructions.
• Task involves predicting a 2D navigation trace, a concrete output for model evaluation.

Limitations

• Description metadata is limited; actual data quality requires manual inspection after download.
• Column-level documentation is absent; field semantics must be inferred after download.
• Row count, file formats, and license information are unknown, which may limit suitability assessment.

Provenance

Source

Voxel51

Freshness

Last updated 2026-05-26 21:35:46; freshness should be verified.