Pick a rollout from the sidebar and watch DreamZero-SO101 imagine the full trajectory from a single starting frame — across three camera views. Use the filters above to isolate one capability at a time.
Send a scene + prompt. Live inference endpoint coming soon.
Open an issuePoint us to an SO-101 episode on HF Hub and we will rollout it.
Use the filter chips above the gallery to isolate one capability at a time, or browse all 35 rollouts together.
Pick a red, blue, yellow, or green cube/lego from a variety of scenes and camera rigs — across four community datasets.
Pick an object and place it in a bowl, bin, or transparent box. Tests trajectory planning beyond just grasping.
Stack the red cube on the blue one — requiring approach, grasp, lift, translate, align, and release in imagination.
Pick all red cubes and place them in the bowl — the model must plan a multi-step sequence for several objects.
Pens, markers, tapes, and scissors — grab each and place into a pen holder. From so101-table-cleanup.
Fold a cloth — the only deformable manipulation task in the training mix, from so101_cloth_folding1.
Six rollouts that start deep inside a trajectory — descend, grasp, and lift phases of held-out episodes.
Eight samples on so101_box_to_bowl_v2 — a dataset the model was never trained on.