Dream Labs has opened Dream API to the public. It takes a robot's camera feed, its current state and a plain-language instruction, and returns motor commands. The Dream-1 model is meant to control arms with grippers without any training data — including hardware it has never seen. The company also raised $90 million in seed funding.
Key takeaways
- Dream-1 was trained on video of people doing everyday tasks, not on a single machine
- Supported robots currently cost between roughly $2,000 and $10,000
- More than 20 hardware partners get a model tuned specifically for them
- Fine-tuning as a service: a small batch of demonstrations, a tailored model within hours
- Today's task range is tabletop work — picking, placing and folding laundry
Video instead of demos on one machine
A record 5 million industrial robots are at work worldwide. Nearly every one is programmed for a single task and cannot transfer a learned skill to another machine. Teaching one robot one task takes specialists months, and new hardware usually means starting from zero.
Dream-1 learns from footage of people rather than demonstrations on a specific unit. That, the company claims, is why its skills transfer across manufacturers. The model predicts what happens next and derives the robot's action from that.
When you get into a rental car, you don't relearn how to drive. You find the controls, and your past driving experience does the rest.
Joel Jang, co-founder and CEO of Dream Labs.
What Dream-1 does not do yet
Zero-shot works today on two-armed robots with parallel grippers. For its 20-plus hardware partners the company tunes the model separately. Models run in data centres across several continents to cut latency. Anyone without a robot can command a real machine in the company's San Francisco office through a browser.
Why it matters
The weight shifts from hardware to model access. If an arm costs as much as a laptop and the intelligence arrives through an API key, the barrier stops being the mechanics budget and becomes the quality of the control policy. The open playground matters more here than the round — it lets you test the model live, mistakes included, rather than watch an edited demo.
What's next
- Support for wheeled mobile robots and dexterous hands is announced, with no date given
- The funding is to go mainly on compute, to widen the range of tasks and supported hardware
Sources
- The Robot Report — Dream Labs raises $90M and launches Dream API
- Dream Labs — Dream-1
