arrowhead/same-goal-different-words-media

Same Goal, Different Words: rollout media

Web media for the project site of Same Goal, Different Words: Spatial Instruction Sensitivity in World-Action Models (CoRL 2026 Workshop: Do Robots Need World Models?).

• Site: https://adeeb10abbas.github.io/same-goal-different-words/
• Evaluation code and released analysis outputs: https://github.com/adeeb10abbas/world-action-model-steerability

One folder per rollout (864 episodes: 3 models × 12 goals × 8 physical starts × 3 wordings), ep/<EPISODE_ID>/:

File	Contents
top.mp4, front.mp4	Executed rollout from the scored camera and the simulator viewport camera (15 fps).
forecast.mp4	The model's generated future next to the executed frames, over the primary request window (all episodes).
dream.mp4	Whole-episode generated track next to the executed one (17 curated episodes; black frames mean no decodable future, not a failure).
first.webp, last.webp, sprite.webp	First and last frames, and a thumbnail sprite used for scrubbing.
timeline.json	Per-tick simulator state used only for post-action scoring and visualization.

Episode ids read RWS-<model>-<scene>-C<start>-<goal>-<wording>, where wording is D (direct), S (target-first) or I (reference-first). These files are web encodes for the site; the scored data and analysis live in the evaluation repository.