---
title: DreamZero Rollout Lab
emoji: 🎬
colorFrom: green
colorTo: blue
sdk: static
app_file: index.html
fullWidth: true
short_description: Ten worlds, WAM predictions, and simulator rollouts
---

# DreamZero Rollout Lab

Explore ten shared bottle-task worlds at Canon DreamZero checkpoints 1000 and
2000. Select a world, scrub the simulator timeline, and inspect any of the
459 independently generated 48-action forecasts beside the observed rollout.
Both the matched camera canvas and the original three-camera simulator strip
are available. All prediction frames and all recorded simulator frames are
included; videos are re-encoded for streaming and interactive seeking.

Higher-FPS simulator recordings publish individually as each replay finishes.
World cards show 10 FPS when upgraded (one frame every three actions); other
worlds retain the original sampling every 60 actions. Each WAM forecast
contains the current frame and eight predicted frames at six-action intervals.
The viewer shows the actual recorded frame step and this sampling difference;
it does not invent simulator frames between observations. Forecasts are generated
independently at each policy call and are never fed back to the simulator.

Evaluation: 22000 actions per world, Blender cameras, 48 actions executed per
call, absolute canonical base-frame EEF20, ABC-source quantile normalization,
canonical grippers and IK to simulator joints. Stats/spec hashes are included
in each world's downloadable JSON. This Space serves saved evaluation artifacts;
it runs no model inference and requires no GPU.

The newest checkpoint with a ready recording is selected by default.
The original 24-GPU checkpoints each have ten complete worlds and score 0/10
task successes; step 1000 has mean maximum reward 0.05 and step 2000 has 0.0.
The resumed 32-GPU checkpoint 3000 is evaluated separately. Its worlds appear
individually as 10-FPS replays finish; the ready count shows the scope of its
displayed results. Checkpoint labels identify the training lineage.

Use the shared URL to link directly to a checkpoint, world and simulator step.
Downloads include the original simulator camera strip, the complete WAM
prediction reel, and the selected world's metrics, timing and provenance.

Video assets are stored in the public [evaluation video repository](https://huggingface.co/datasets/OneWorld-AI/dreamzero-canon-sim-eval-videos); the Space hosts the viewer and metadata. Refresh the page to load newly published recordings.