Transforming an ordinary machine into an intelligent automaton no longer demands exorbitant budgets or months of custom programming. For a mere $3 per hour—or just 75 cents for a slightly streamlined physical AI variation—developers can now tap into advanced robotics intelligence. Dream Labs, a San Francisco enterprise launched earlier this year by former pioneers of Nvidia’s GR00T robotics division, has officially emerged from stealth with a staggering $90 million seed funding round alongside the debut of the Dream API, touted as the market's pioneer public API for a robotic foundation model.

The mechanism functions almost like sorcery: engineers feed the API a live camera stream from their mechanical device alongside straightforward text instructions, enabling the software to pilot the apparatus seamlessly, even if it has never previously encountered that specific hardware build. According to the enterprise's leadership, this breakthrough represents the long-awaited missing jigsaw piece in physical artificial intelligence.

The hardware is ready. What’s been missing is a brain any robot can use. 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. Dream-1 approaches a new robot the same way, so a new robot can do useful work in minutes, not months. — CEO and co-founder Joel Jang

Unlocking Cross-Embodiment Intelligence

At the heart of the platform sits Dream-1, an AI architecture schooled extensively on footage depicting humans executing daily chores—a methodology mirrored by competitors like Figure with Index and Germany's Neura Robotics. Rather than learning from a single machine tied to rigid software instructions, the model analyzes video to forecast subsequent visual actions before translating those predictions into physical commands. This video-centric training enables skills to transfer fluidly across disparate hardware ecosystems.

The company maintains that out of the box, Dream-1 can instantly pilot any machine featuring dual arms and parallel grippers without prior calibration—a feat roboticists describe as zero-shot capability. Historically, cross-embodiment transfer—teaching a secondary robot to replicate tasks mastered by a primary machine—has plagued the industry. Because nearly all of the approximately 5 million industrial robots active globally are custom-programmed for distinct bodies, differing arm geometries, friction levels, and motor mechanics, insights gained by one unit rarely benefit another.

Just half a year ago, adapting internal models to unfamiliar hardware required thirty minutes of dedicated machine data, a timeline that once stretched to 100,000 engineering hours. As CEO Joel Jang notes, that timeline has collapsed entirely, reducing the requirement simply to having the robot present.

Reframing Control as a Video Challenge

The core innovation lies in shifting the paradigm of mechanical control entirely into the realm of video forecasting. By treating movement as a predictive visual puzzle, the software bypasses traditional integration hurdles. While zero-shot pick-and-place operations achieve an 80% success rate and secondary tasks hover between 50% and 60%—metrics currently insufficient for rigorous factory assembly lines—Dream Labs has deliberately set its sights elsewhere.

Rather than chasing the 99% accuracy thresholds demanded by specialized industrial manufacturing, the startup focuses on delivering everyday competencies for versatile, general-purpose machines. For deployments requiring commercial-grade precision, the firm provides a rapid fine-tuning service, transforming client demonstration data into tailored models within a matter of hours.

  • Flagship flagship model available at $3 per hour
  • Budget flash model available at $0.75 per hour
  • Zero-shot compatibility for dual-arm systems with parallel grippers
  • Support for more than 20 hardware manufacturing partners

This rapid turnaround dismantles the traditional paradigm where deploying a $5,000 to $10,000 robot required reserving GPUs, hosting open-source codebases like GR00T or π, and enduring months of custom integration. Furthermore, it empowers enterprises to deploy heterogeneous fleets—mixing and matching hardware from diverse manufacturers without doubling or tripling ongoing maintenance and training expenditures.

Backing and Market Competition

The substantial $90 million seed financing was spearheaded by B Capital, alongside contributions from A.Capital, Kindred Ventures, Unusual Ventures, SV Angel, Wing VC, Scribble VC, and Adrian Aoun. Investors emphasize that as hardware costs plummet to match high-end laptops, intelligence has become the primary bottleneck.

Operating expenses shift traditional cost models toward an hourly utility structure. Running a flagship unit continuously around the clock totals roughly $26,000 annually, while the economy model drops expenses to approximately $6,600. Given that most machines operate intermittently, and historical trends indicate continuous downward pressure on token pricing, these operational expenses remain highly competitive.

Nevertheless, the competitive landscape remains intense. Well-funded rivals include Skild AI with its robot-agnostic Skild Brain, Physical Intelligence and its π models, Google's Gemini Robotics, and Nvidia's GR00T platform. To stand out, Dream Labs relies heavily on friction-free distribution: instant API key acquisition, pay-as-you-go pricing, and transparent testing environments like the Robot Playground, which allows prospective clients to issue live, unedited commands to real hardware housed in San Francisco.

With wheeled mobile architectures and dexterous hands next on the roadmap, the newly acquired capital will primarily fund the massive compute infrastructure required to scale Dream-1's capabilities across an expanding array of mechanical platforms.