live pointillism of a real DROID episode vs its counterfactual. Orange where the two futures disagree.
From a single real clip, we generate infinite alternate paths for your robot, and measure every one against the original. Proven different, never just plausible.
Every clip is rendered next to its unchanged replay from the same starting frame and seed, then measured. Clips that fail any check are discarded before delivery. Nothing failed is ever billed.
One real episode, two futures. The strip tracks commanded joint-space divergence against rendered pixel divergence. The render follows the coordinates, not the text prompt.
A real DROID episode of a towel fold, and a counterfactual that folds from the other side. Both futures render from the same starting frame with the same seed. The only input that changed is the commanded trajectory.
Rendered motion stays above the check floor in all seven segments of each future. Divergence between them reaches 36 gray-levels per pixel by the final second, and the whole pair was staged and rendered in 20 minutes from a public dataset.
“Under counterfactual instructions, models hallucinate absent objects or change object attributes.”
A real teleop episode costs $15 to $40 and yields one trajectory. Each of your episodes can yield dozens of verified variants without another hour of teleop.
A world model is fine-tuned on your real episodes: your embodiment, your cameras, your dynamics. No CAD assets. No simulator gap.
Each clip branches from a real frame of a real episode with a modified command track. A different reach. A slower approach. A slip that never happened.
Divergence proves the change took effect. Motion checks catch hallucinated stillness. Fails are discarded before you ever see them.
Ranked by what verification protects you from.
Near-misses and slips, generated from your own data and verified physically real, so your policy learns the edge cases before a real robot finds them.
Paired real-vs-counterfactual clips hold the scene fixed and vary only the command. The measurement shows the alternate action produced a genuinely different outcome.
Trajectory diversity a single demo physically can't contain. Every generated clip traces back to its source episode, so augmentation helps your policy instead of quietly poisoning it.
Run checkpoints against a fixed suite of verified scenarios and compare divergence telemetry before burning real robot hours.
World models drift. Long rollouts lose texture fidelity, and object interactions beyond the branch window are frontier work. We publish those limits with the same instruments we use to verify your clips, and the receipt tells you exactly what each clip is good for. That honesty is the product.
Send a sample of your real episodes. You get back the multiplied set in LeRobot format, every clip carrying its provenance and verification receipt, plus a one-page report. Success is measured one way: your policy, trained on real plus ours, beats real-only on your own eval.
worlds.alakazam.gg · built on NVIDIA Cosmos world models · your data stays yours