Dream Labs
VerifiedTrack: raisedWebsite ↗𝕏 @dreamlabs_aiLinkedIn ↗
potential 80 · scoop 58
Dream Labs sells a public API for zero-shot robot control priced by the hour. Four ex-NVIDIA GEAR Lab researchers raised a $90M seed after releasing DreamZero in February 2026.
Scores run 0 to 100. Potential: how likely the firm is to become a real winner, judged across the six dimension scores and held consistent with them, whether or not anyone has heard of it. Scoop: how unlikely a well-read tech investor is to have heard of it yet. High means it has not reached the outlets that reader follows, whatever the firm's quality or age. How our top potential scores did.
Four NVIDIA GEAR Lab researchers released DreamZero in February 2026 and raised $90M six months later to ship the first public API for cross-hardware robot control.
Snapshot
- What they build
- Builds a foundation model that trains robots from video and natural-language instructions
- Who buys
- Robotics developers and companies deploying physical AI
- Comparables
- OpenAI for robotics · Figure/1X-style model layer
- Sub-industry
- Robot foundation model / AI for robotics training
Backed by top VCs shortly after founding
How we found them
- Surfaced in issue
- Oct 26, 2026
- Announced
- Oct 20, 2026 (seed funding)
- Founded
- 2026
- Round
- seed · $90,000,000
- Investors
- B Capital, Kindred Ventures, A.Capital, Unusual Ventures, Wing VC, SV Angel, Scribble VC
- HQ
- San Francisco, CA, US
- Discovery track
- raised
Dimension scores
- Team: 95 out of 100
- Product: 80 out of 100
- Traction: 65 out of 100
- Market: 85 out of 100
- Moat: 50 out of 100
- Capital: 90 out of 100
Analysis
The four founders, Joel Jang, Yunhao Ge, Scott Reed, and You Liang Tan, all worked together at NVIDIA GEAR Lab on the GR00T project. They left in April 2026 after six months of internal tests on cross-embodiment transfer.
The team released DreamZero in February while still at NVIDIA. The public API launched on October 19 with per-hour pricing at $3 for Dream-1. It logged 372k inference calls across 31 unseen robot types. An independent Robocurve eval scored it at 59.3 percent, roughly 16 times cheaper than frontier VLAs.
The models combine video prediction with action generation from egocentric video. This removes the need for months of specialist data collection per new arm. Research-beta terms still block commercial use and close-proximity tasks, so paid production deployments remain ahead.
What they do
Dream Labs provides a public API (Dream API) and fine-tuning service that lets developers connect bimanual robot arms with parallel grippers to a World Action Model (Dream-1) for zero-shot control from camera feeds and natural-language instructions, without robot-specific training data.
Why now
Hardware arms dropped from “luxury car” to “high-end laptop” pricing in recent years, creating demand for transferable intelligence; 2026 maturation of video world models (DreamZero and Cosmos releases) enabled the zero-shot WAM results that broke the data-per-robot bottleneck; post-ChatGPT developer appetite for simple APIs; compute scaling (GB200-class) making real-time video diffusion viable for control.
Founding story
The four founders met and collaborated at NVIDIA’s GEAR Lab on GR00T robotics research. In February 2026 (still at NVIDIA), they released DreamZero and began using the term “World Action Model” (WAM) for video-prediction-plus-action models that learn from human egocentric video for generalization. After ~6 months of internal validation (including cross-embodiment transfer with minimal robot data), they spun out to found Dream Labs (~April 2026) to productize a public API.
Competitive position
Dream Labs occupies a distinct lane: hardware-agnostic public API with explicit per-hour pricing, zero-shot cross-embodiment emphasis via WAMs (vs. mostly imitation or single-embodiment), and fine-tuning service, positioned as “the brain, not the body” for cheap existing arms rather than end-to-end robot sales or closed proprietary stacks.
Defensibility
Genuine moat is thin and execution-dependent: the team’s first-mover status in publicly shipping a WAM-based cross-embodiment API (coining the category) plus fine-tuning service creates switching costs and brand; shared video representation theoretically scales without per-robot datasets, but video data is widely available and architectures are reproducible; no proprietary data lock-in, regulatory license, or exclusive talent beyond the founding four (who defined key priors at NVIDIA).
Network density
4 of 4 founders from the same NVIDIA GEAR Lab / GR00T team. The broader early team (~12 researchers and engineers total) shows high concentration from this group and related robotics labs (e.g., Berkeley RAIL, Open Robotics alumni via Tan).
Early team signals
Notable investors/advisors: B Capital (led; GP Daisy Cai), Kindred Ventures (co-led; founder Steve Jang joins board), A.Capital, Unusual Ventures, Wing VC, SV Angel, Scribble VC; angels including Thomas Wolf (Hugging Face CTO), Adrian Aoun. Hardware partnerships with >20 unnamed vendors (examples referenced: Unitree, Robotis).
Traction
- Public Dream API and browser playground live since Oct 19, 2026 launch
- >20 hardware partners (unnamed specifics)
- Independent Robocurve RoboDojo-RC Tier 1 eval showing 59.3% mean score (first >50%, outperforming Gemini 4 Argon/Claude Opus 5.5/GPT-6 variants at ~16x lower cost and 1.7x faster)
- X-reported beta usage (31 unseen robot types, 53 countries, 372k+ inference calls shortly after launch)
- Press coverage across Forbes, SiliconANGLE, The Robot Report, Kindred Ventures announcement
- Small but credible team execution on launch
- No named revenue customers or ARR yet
Opportunities
- Hardware price collapse continuing to flood market with cheap arms needing brains
- Potential EU AI Act or similar robotics standards favoring auditable/evaluable models
- Expansion of video foundation model capabilities (e.g., longer context, better simulation) directly unlocking more tasks/embodiments
- Developer ecosystem pull via public API + playground (similar to early LLM APIs)
- Partnerships with low-cost arm makers (20+ already referenced) for bundled intelligence
Risks / red flags
- Research-beta status with restrictive API terms (commercial use and human-proximity tasks require separate agreements) limits near-term revenue and real-world deployment signals
- All four founders from one prior organization (NVIDIA GR00T) signals potential lack of operational diversity under scaling pressure
- $90M seed is outsized and implies high valuation with limited public traction beyond beta usage metrics and one independent eval (Robocurve)
- Compute-heavy model with funding earmarked primarily for inference/training capacity creates runway dependency
- Performance claims (e.g., Pareto frontier on RoboDojo-RC) rely on new or company-adjacent evals with no long-term named paying customers or production deployments disclosed
- Early-stage regulatory/safety exposure for physical robots (no visible HIPAA/ITAR-style prep mentioned)
Hidden bets
- World Action Models (WAMs): joint video prediction + action generation from pretrained video diffusion backbones (Cosmos/GR00T lineage)
- Research-beta API terms restrict commercial/paid work and close-proximity human tasks
- Inference priced per hour ($3 Dream-1 / $2 flash)
- Fine-tuning-as-a-service for adaptation
- Heavy compute spend (funding primarily for this)
- Focus initially on tabletop bimanual grippers before mobile/dexterous hands
Key relationships
- B Capital (led)
- Kindred Ventures (co-led)
- A.Capital
- Unusual Ventures
- Wing VC
- SV Angel
- Scribble VC
- Thomas Wolf (Hugging Face CTO)
- Adrian Aoun
Founders
Citations
- https://dreamlabs.ai/
- https://dreamlabs.ai/research/dream-1/
- https://x.com/dreamlabs_ai/status/2112230800926331293
- https://pr.ai/threads/dream-labs-robotic-foundation-models-san-francisco-california-usa.28405/
- https://roboticaguide.com/dream-labs-dream-1-price-robot-api/
- https://kindredventures.com/announcement/dream-labs-building-the-brain-for-everyday-robots/
- https://joeljang.github.io/
- https://www.mt.co.kr/en/future/2026/10/20/2026102009494495865
- https://gyhandy.github.io/
- https://eu.36kr.com/zh/p/4033427959534470
- https://youliangtan.github.io/
- https://36kr.com/p/4033427959534470
- https://x.com/jang_yoel/status/2112232332333900170
- https://x.com/dreamlabs_ai
- https://www.youtube.com/watch?v=fqkp_wkov6M
- https://github.com/octo-models
- https://octo-models.github.io/
- https://dreamlabs.ai/careers/
- https://www.forbes.com/sites/johnkoetsier/2026/10/19/robot-brains-for-75-centshour-former-nvidia-researchers-raise-90-million/
- https://siliconangle.com/2026/10/19/dream-labs-raises-90m-to-help-build-more-intelligent-robots-with-simple-video-feeds/
- https://futuretweets.com/gemini-robotics-vs-nvidia-groot-vs-figure-helix-2026/
- https://compareais.net/blog/posts/embodied-ai-companies-2026.html
- https://venturecapitaltracker.com/dream-labs-90m-seed-robot-foundation-models
- https://www.blackscarab.ai/news/dream-labs-90m-seed-robot-intelligence