Dream Labs Launches Dream-1 Robot Policy with $90M Seed

Dream Labs is emerging from stealth mode with $90 million in seed funding. The company is launching Dream-1, a manipulation policy designed specifically for bimanual robots equipped with parallel grippers. The model is available to users worldwide through a public API. Dream-1 can control unseen robots for unseen tasks in unseen environments out of the box. No machine learning background is required to get started with the system.

Dream-1 can zero-shot generalize to many bimanual, parallel-gripper robots. It can control the original PR2 robot without using embodiment-specific data. Robocurve independently evaluated Dream-1 ahead of its launch. The model scores at 16x lower cost per task attempt than Gemini 4 Argon. It is also 1.7x faster than GPT-6 Astra. Running Dream-1 costs $3 per hour, while Dream-1-flash costs $2 per hour. Each call to the model returns 1.6 seconds of actions. An hour of usage equals approximately 2,250 calls, resulting in a cost per action of about $0.0013.

Fine-tuning Dream-1 for a specific robot using FaaS requires collecting data on the user's own robot. Fine-tuning on YAM with different end effectors took 98 minutes of data. Across twelve robots from FaaS beta partners, fine-tuning lifted the language-following score on ten of them. Dream-1 currently has real limits and is described as a starting point for general robot foundation models.