- AngelsRound
- Posts
- Saturn Dynamics đȘ
Saturn Dynamics đȘ
AI simulation for robotics.
Spotlight
What if a robot could make its mistakes before entering the real world?
Quick Pitch: Saturn Dynamics is building Atlas, an AI simulation where robots can train and test before entering the real world.
The Problem
Real World Learning Is Expensive: Every failed attempt consumes hardware time, supervision, and costly data collection.
Simulation Has Limits: Traditional physics engines can struggle to generalize across changing real world environments.
World Models Are Compute Heavy: AI simulations, known as world models, can require too much compute to run directly on robots.
Why It Matters
Physical AI and embodied robotics are projected to grow from roughly $21 billion in 2025 to more than $180 billion by 2034.
As more AI moves into robots, vehicles, and factories, those systems will need places to train and test before operating in the real world. World models could become a key infrastructure layer underneath that growth.
Snapshot
Industry: Physical AI / World Models
Headquarters: London, UK
Founded: 2025
Backed By: Alumni Ventures
Funding: Raised $1M (Seed)
Traction: Atlas shipped in April 2026 with a five person team
Target Customers: Robotics companies and industrial OEMs
Founderâs Edge
Elisa Seghetti: Previously built an AI software agency from 3 to 35+ engineers and leads Saturnâs commercial strategy.
Simone Totaro: Reinforcement learning researcher with experience across robotics, drones, Meta Reality Labs, and Mila.
Cesc Cunillera: Physics PhD whose research combining simulations with deep learning informs Saturnâs physics based approach.
Playing Field
NVIDIA Cosmos: A leading world model platform, but more compute intensive in Saturnâs benchmarks.
Video Based World Models: Many start with large video models and adapt them for robotics, which can add size and compute requirements.
Saturn Dynamicsâ Edge: Purpose built for robotics, with smaller models designed to run directly on a robotâs hardware.
Analysis
Bulls Case đ
World models could become core infrastructure for physical AI.
Saturnâs efficiency could make them accessible to a much larger market.
Rare mix of physicists, RL researchers, and repeat founders.
Running on the robot could turn compute efficiency into a distribution advantage.
Bears Case đ
Performance claims still need independent validation.
NVIDIA owns the hardware, the ecosystem, and a competing product.
Larger competitors could replicate Saturnâs efficiency gains.
Atlas currently focuses on navigation, with harder robotics tasks still ahead.
Verdict
Saturnâs real opportunity is not making simulation cheaper. It is making simulation local.
If robots can simulate and evaluate actions on their own hardware, world models could move from a training tool to part of the robot itself. That would make efficiency more than a cost advantage. It could determine who owns a new layer of the robotics stack.
The Startup Pulse
Another happening week in startup funding. Three signals from this week:
AI agents are moving from answering questions to actually completing work
Physical AI is creating new markets around chips, networking and engineering
Public markets are bringing more discipline to private market valuations
Instinct: Raised $1B at a $10B valuation. Personal AI is moving beyond answers toward actually booking, buying, and transacting for users.
EliseAI: Landed $350M at a $4B valuation. Its AI already touches roughly one in six U.S. apartments, a sign of how quickly vertical AI can scale when embedded into everyday workflows.
Armadin: Secured $255.5M at a $2.5B+ valuation. The former Mandiant CEO is now betting autonomous agents can continuously find vulnerabilities before attackers do.
Subscribe on Youtube
Strategic breakdowns of how the biggest AI companies actually work â their business models, their moats, their vulnerabilities. This week: Why A.I. companies are quietly building inside Iceland's volcanoes.
Written by Ashher
