Helix 2.5 Walked Into 30 Houses It Had Never Seen

Figure’s Helix 2.5 paper is the most useful robotics result I have read this year, and it is not because the robot looks human. It is because the training recipe is finally starting to resemble language models: pretrain on a huge pile of human behavior, then ask the policy to work in a room it has never entered.

The company ran three long-horizon chores — tidy a living room, fold towels, make a bed — across 30 real homes with no data collection and no adaptation in those spaces. Success meant finishing the whole task. No partial credit. The Index-pretrained policy hit 56%. A policy trained from scratch hit 9%.

The scaling law is the story

Figure says Index, their human-behavior dataset, now adds roughly 35 minutes of new experience every second, and they have committed $3.5B of compute to Helix. As Index grew, zero-shot transfer improved smoothly enough to plot. That is the same curve we already believe for text and images. Seeing it on whole-body control is new.

• Pretraining on human video and teleop beats collecting only in the target warehouse.
• Generalization is measured as finished chores, not pretty rollouts.
• Simulation still matters for stairs and terrain, but the home test is the headline.

What app and ML people should take from it

If you build on-device models or small chatbots, Helix is a reminder that data diversity beats scene-specific overfitting. A focus app that only ever saw one user’s habits will feel brittle. A retrieval corpus that only covers your own docs will fail the first real customer. The 6× gap between scratch and pretrained is the same gap I see when a chatbot has memory versus a chatbot that only has the current message.

Zero-shot is not magic. It is the bill you already paid in data, showing up in a room you did not rehearse.