WORLD MODELS / EMBODIED DATA

Interaction data for
world models.

Observations, actions and next observations. Reproducible collection, validation and custom delivery for embodied AI teams.

Download 3 sample episodes ↧Discuss a data request →
12,000episodes
818,161transitions
3simulation subsets

WM Interaction v1 · simulation records · fixed-seed random discrete-action policy. Dataset and model evidence have separate scopes.

01 / PUBLIC RELEASE

WM Interaction v1

Read the dataset card ↗

Three published subsets. Episode counts, transition counts and formats are derived from the release manifests.

Toy · Easy

Episodes
5,000
Transitions
160,000
Observation
8 × 8 × 3
Discrete actions
4
Mean length
32

Toy · Hard

Episodes
5,000
Transitions
320,000
Observation
16 × 16 × 3
Discrete actions
4
Mean length
64

Crafter

Episodes
2,000
Transitions
338,161
Observation
64 × 64 × 3
Discrete actions
17
Mean length
169.0805

Crafter uses the canonical-object-order variant documented in the dataset card. The two Toy subsets use obstacle-navigation environments. All v1 actions are discrete.

Download full v1 · 822 MB ZIP ↧File checksum manifest ↗

ZIP · 822,009,439 bytes · SHA256 63d7b99196b34f8c3611a4a34fb57d159aa55cbcbb2022f4b9c0858878e095d0

02 / INSPECT THE RECORDS

Follow an interaction. Step by step.

Episode 0 from each subset, 255 recorded transitions in total. Move through time to inspect the original observations, actions and rewards.

Recorded observation before the action
observation
Recorded observation after the action
next_observation
0 / 31
Subset
toy
seed
0
Action ID · source label
1 · move_down
Observation shape
8 × 8 × 3
reward
-0.01
done
false
truncated
false

Loading the episode timeline…

NPZ SHA256 8422a7390ea91efa2fbdedafa0e9cd61154981eb823d79619a12a2b2665ff0ff

Preview PNGs preserve the original RGB pixels; display enlargement uses pixelated rendering. Action labels come from the source environment. Changes can also reflect environment dynamics. Sample provenance ↗

Download the original samples ↧Read with NumPy ↓

Sample ZIP · 147,749 bytes · SHA256 7a8bcfb82deb5adf50ccd0e16fb4e1a3dc7138b601eb9583570d5d9861385862

03 / DELIVERY CONTENTS

Data with a delivery record.

Transition records

One NPZ per episode: obs, action, next_obs, reward, done and truncated. Images are uint8 HWC; actions use environment-specific discrete IDs.

Metadata and file hashes

meta.jsonl records environment, seed, length, observation shape and action count. SHA256SUMS verifies the delivery files.

Order and reproduction

Custom orders include ORDER.yaml and REPORT.md with the original parameters, actual output counts and a reproduction command.

Optional action diagnostics

Request action_sensitivity.json for action-sensitivity distributions, pairwise next-observation differences and low-score episode review.

04 / VALIDATION

Checks you can inspect.

12,000 / 12,000

Independent collection matches

Every episode NPZ matches by SHA256 across two complete collections. Runtime versions and source fingerprints accompany v1.

818,161

Transitions replayed

All recorded transitions were replayed, checking observations, next observations, rewards and termination flags.

16,000

Crafter counterfactual timepoints

All toy timepoints were analyzed. Crafter uses 8 points per episode: 4.7315% coverage, separately from full transition replay.

Byte reproduction requires the same environment implementation, configuration, seed and pinned Python, NumPy and zlib versions. Changed dependencies require another hash acceptance run. Action-sensitivity scores depend on image resolution and are not rankings across environments. Full validation scope ↗

05 / CUSTOM DELIVERY

Start with the data specification.

Define the environment, observation and action contract, volume and acceptance criteria. We will confirm scope, schedule and quotation.

Open email ↗

Prepare a draft, then send it through your email client.

Standard collection supports Toy and Crafter simulation data. Customer environments, continuous actions, real-robot recordings and additional formats require scope and acceptance review. Transition totals depend on actual episode lengths.

06 / READ AND REPRODUCE

Inspect the files with NumPy.

After downloading the sample ZIP

python -m zipfile -e wm-interaction-v1-samples.zip wm-samples
python -m pip install numpy
python wm-samples/read_sample.py wm-samples toy

Python 3.12 and NumPy. The reader checks the episode SHA256 before loading. Replace toy with toy_hard or crafter to inspect the other subsets.

A verified custom-order example

env: toy
episodes: 8
difficulty: medium
seed: 42
obs_size: 16
action_subset: [0, 1, 2, 3]
include_action_sensitivity: true

This order generated 8 episodes and 384 transitions. The same parameters appear in the delivery report. Download YAML ↧

Licensing and attribution: KineJing contributions are CC BY 4.0 where rights are held. Crafter assets retain upstream MIT terms. Dataset cards and third-party notices define the complete terms. Training benefits and transfer to real robots have not been measured for this release. Data license ↗ · Third-party notices ↗