FULL STORY

Dream Labs Launches Dream-1, Claiming a New Robotics Frontier

Dream Labs, founded by ex-NVIDIA researchers, emerged from stealth with robotics foundation model Dream-1, backed by a $90M seed round. Independent pre-launch tests and official claims point to a new performance frontier for robot control.

2026-10-20 ~ 2026-10-21 · 3 episodes · 26 posts

Episode 1 · Robocurve's Pre-Release Eval: Dream-1 Tops RoboDojo, All Raw Data Released (2026-10-20, 6 posts)

Robocurve independently evaluated Dream Labs' Dream-1 robot-control model before its release, concluding it sets a new Pareto frontier: 16x cheaper and 1.7x faster than the prior best while performing better. It then released all raw run data for community reproduction and audit.

Confirmed

  • Dream-1 topped the RoboDojo-RC Tier 1 real-robot leaderboard at 59.3%, the first model to break 50%, well ahead of Gemini-family models (Gemini 4 cited as comparison).
  • Per-run cost is about $0.15.
  • Robocurve published all 120 Dream-1 runs, including full transcripts, logs and videos, with a filterable run browser and CSV download.
  • The site also opened its full evaluation database of 2,469 runs, filterable by benchmark (RoboDojo-RC Tier 1, RoboHarm, etc.).

Why it matters

  • The evaluation predates the model's release, making it a third-party assessment rather than vendor marketing, which adds credibility.
  • An order-of-magnitude cost drop plus a speed gain, if reproduced, would substantially lower deployment barriers for real-world robotic tasks and reshape the cost-performance landscape of robot foundation models.
  • Full raw data release (transcripts, logs, videos, CSV) sets a transparency benchmark for robot foundation model evaluation.

Episode 2 · Ex-NVIDIA GR00T team launches Dream Labs with Dream-1 robot foundation model and open API (2026-10-20, 18 posts)

Dream Labs, founded by four core members of NVIDIA's GR00T team (Joel Jang, Ge Yunhao, Scott Reed, You Liang Tan) who left NVIDIA around February 2026, has emerged from stealth with Dream-1, a robot foundation model opened to all developers via a public API starting at $3/hour. The company also announced a $90M seed round, placing it in the top 1% of US enterprise-software seed rounds. It claims to be the first public robotics API.

Confirmed

  • The four founders left NVIDIA around February 2026; Scott Reed was previously a Google DeepMind researcher; the company emerged from stealth about six months after founding
  • Dream-1 and the Dream API opened to everyone on launch day at $3/hour; the model is claimed to zero-shot control unseen dual-arm robots (parallel gripper robots per QbitAI), perform never-demonstrated tasks, and operate in never-visited locations
  • Zero-shot claims detailed further: controlling 27 never-seen robots in simulation and on real arms, including the 18-year-old PR2, with no training data for that robot; the team also claims roughly 1/16 the cost of traditional approaches
  • Technical approach: a "World Action Model" — first imagining video, then acting — learning how the world moves from videos of humans performing tasks
  • Funding: $90M seed led by B Capital with Kindred Ventures co-leading, 500 Global and others participating; the goal is a general foundation model controlling robots across vendors (note: one post by @lukasmziegler says $9M, inconsistent with his own and other sources; the majority-consistent $90M is used)
  • Market context: the company estimates over 5 million industrial robots in factories and warehouses worldwide, per @emmanuelvivier
  • Investor and advisor George Harik (ex-OpenAI LP) highlights zero-shot generalization to new embodiments without targeted training
  • Motivation: even three years after the RT-X era, using foundation models on one's own robots still requires data collection and fine-tuning; cross-task, cross-embodiment generalization remains the pain point

Unconfirmed

  • Figures such as "27 unseen robots zero-shot" and "1/16 cost" come from reposts of official materials without independent verification

Why it matters

  • If zero-shot cross-embodiment capability holds, developers could deploy on new hardware without data collection or fine-tuning, sharply lowering the barrier for robotics applications
  • Hourly-billed public APIs turn robot control into an on-demand service, mirroring the foundation-model API business model
  • The $90M seed size signals how heavily capital is betting on robot foundation models

Episode 3 · Dreamlabs Releases Dream-1, Claiming Faster Speed and Higher Scores Than GPT-6 (2026-10-20, 2 posts)

Dreamlabs AI released Dream-1, claiming scores 1.3x and speed 1.7x that of GPT-6 Astra Ultrafast; third-party benchmarks showed it completing 41% of subtasks within 5 minutes.