> ## Content Index
> Fetch the complete content index at: https://globalfeed.ai/llms.txt
> Use this file to discover other available public pages before exploring further.

# Vsim says its simulator trains robot skills in minutes
- URL: https://globalfeed.ai/en/vsim-says-its-simulator-trains-robot-skills-in-minutes/
- Published: 2026-09-18T16:52:59.000Z
- Updated: 2026-09-18T16:52:59.000Z
- Description: The Cambridge start-up runs millions of virtual attempts to train robot skills in minutes, while Nvidia scales the same approach with AI agents.
- Author: GlobalFeed Editor
- Tags: Vsim, NVIDIA, robotics, x-nvidia, dil-en, denetim-gerekli, video

At Vsim, a British start-up based in Cambridge, a robot called Freddo walks across the office, recognises a plastic bottle handed to it and grasps it. The company says the skills took only a few minutes to train and upload, and that rival systems could need days for the same result.

The method is virtual simulation: a task is repeated millions of times in software until an optimal solution, known as a policy, is found, then transferred to the hardware. This is a standard approach in robotics. Nvidia's Isaac Sim works the same way, and Vsim founders Michelle Lu and Kier Storey both worked on an early version of it.

## Algorithms rebuilt for GPUs

The pair founded Vsim in 2022 to build their own training environment. Storey says the algorithms behind most robotic simulations date back to the 1970s and 1980s and are a poor fit for GPUs. Starting from scratch let them optimise for those chips. Eighteen months in, Lu says, they had a fully functional, very high performance simulator.

The software is light enough to run on the hardware Freddo carries, so the robot can run tens of thousands of simulations while moving. Storey says it can look about a second ahead across 20,000 possible combinations of events. Lu notes that humans, animals or other robots can force a sudden change of strategy, and the robot has to adapt fast to stay safe and on mission.

## Scale and limits

Vsim has 10 engineers. Nvidia does not build robots but runs a robotics software division with hundreds of engineers, including a world model called Cosmos that gives robots an understanding of real world physics. Spencer Huang, director of product for robotics at Nvidia, says grabbing a bottle is manageable, but long horizon tasks such as taking a bottle, filling it and pouring are the hard part. This year Nvidia began using AI agents to build virtual environments and validate training results.

Rika Antonova, associate professor at the University of Cambridge, works with MuJoCo, the open-source system owned by Google DeepMind. She calls Vsim's very fast simulation promising, but says simulated environments remain rough approximations, with deformable objects and cutting still hard to model. A second Vsim robot, Nacho, is due to join development.

Full report at [BBC News](https://www.bbc.co.uk/news/articles/c79g0j3d4q9o?at%5Fmedium=RSS&at%5Fcampaign=rss&ref=globalfeed.ai).