HiggsBoson/vizdoom-llm-inverse-dynamics-benchmark

VizDoom LLM Inverse Dynamics Benchmark

Dataset Summary

This dataset is a small, manually inspectable benchmark for evaluating how well large language models and vision-language models can act as inverse dynamics models in a first-person game environment.

Each example contains a 9-frame temporal window centered on a labeled timestep t:

[x_(t-4), x_(t-3), x_(t-2), x_(t-1), x_t, x_(t+1), x_(t+2), x_(t+3), x_(t+4)] -> a_t

where a_t is the action taken at time t.

The benchmark is intended for evaluation, not large-scale training. It was designed to support questions like:

• Can an LLM or VLM infer the player action from short-term visual dynamics?
• Which actions are easiest or hardest to distinguish?
• How does model performance vary with human-annotated difficulty?

The current public release contains:

• 1 split: benchmark
• 100 examples
• 9 image columns per example for the temporal window
• action labels from a small discrete navigation action space
• optional manual annotation fields for difficulty and difficulty rationale

Task Framing

This benchmark evaluates inverse dynamics prediction from egocentric visual observations in VizDoom. The goal is to predict the discrete action taken at the center frame.

Supported action labels:

• forward
• turn_left
• turn_right
• strafe_left
• strafe_right
• idle

Data Source and Construction

The benchmark was constructed from trajectories collected in VizDoom using a navigation-oriented action space. Each benchmark row is derived from a labeled decision timestep and exports:

• the center-frame action label
• 4 frames before the decision
• the frame at the decision timestep
• 4 frames after the decision
• a contact-sheet image for quick human inspection

The benchmark was intentionally stored in a format that is easy to:

• inspect manually
• annotate in a spreadsheet
• export into Hugging Face Image features

Data Fields

Each row includes the following fields:

• sample_id: unique sample identifier
• sequence_image: a contact sheet showing all 9 frames side by side
• image_before: alias for frame_t
• image_after: alias for frame_t_plus_1
• frame_t_minus_4
• frame_t_minus_3
• frame_t_minus_2
• frame_t_minus_1
• frame_t
• frame_t_plus_1
• frame_t_plus_2
• frame_t_plus_3
• frame_t_plus_4
• action: ground-truth action name
• ground_truth_action: duplicate of action for compatibility with older evaluation flows
• action_id: integer action id
• difficulty: optional human annotation of perceived difficulty
• difficulty_reason: optional short human explanation for the difficulty label
• episode_id: source episode identifier
• split: split name
• label_frame_index: timestep index of the labeled action within the episode

Intended Uses

This dataset is intended for:

• benchmarking LLMs and VLMs on inverse dynamics prediction
• prompting studies over short visual sequences
• error analysis across action types
• analysis conditioned on human difficulty labels

Reasonable evaluation setups include:

• prompting with the full 9-frame context
• using only image_before and image_after as a two-frame baseline
• comparing temporal prompting versus single-frame prompting

Out-of-Scope Uses

This dataset is not intended for:

• training a robust general-purpose policy
• estimating real-world human behavior
• evaluating open-world navigation competence beyond this narrow benchmark setup
• drawing broad conclusions about embodied reasoning from only 100 examples

Annotation Notes

The difficulty and difficulty_reason fields are manual annotation fields. Depending on the current uploaded version, some or all of these fields may be blank. Blank values should be interpreted as not yet annotated, not as an explicit difficulty judgment.

Limitations

• The dataset is small and intended for evaluation rather than training.
• It comes from a single game domain and a narrow action space.
• Visual ambiguity can arise from motion blur, repeated textures, and weak frame-to-frame changes.
• Human difficulty labels are subjective and may evolve across versions.

Loading the Dataset

from datasets import load_dataset

dataset = load_dataset(
    "HiggsBoson/vizdoom-llm-inverse-dynamics-benchmark",
    split="benchmark",
)

print(dataset)
print(dataset.column_names)
print(dataset[0]["action"])

Suggested Evaluation Prompt

One natural evaluation prompt is:

You are given a short sequence of first-person game frames centered on a decision timestep. Predict the single action taken at the center frame from: forward, turn_left, turn_right, strafe_left, strafe_right, idle.

This dataset is also compatible with pairwise prompting using only image_before and image_after.

Acknowledgements

This benchmark was built from a VizDoom inverse-dynamics data pipeline and is intended to support research and course-project style evaluation of LLM/VLM capabilities on action inference from short visual sequences.