World Labs presented a simulation engine for training robot controllers entirely in virtual environments. According to The Decoder, from one real-world task, the system creates thousands of controllable variants.
The trained models then ran for one hour on five different robotic platforms without human intervention. This indicates testing the approach on multiple types of hardware, but does not prove readiness for wide-scale deployment.
The practical value of the technology depends on how well virtual training transfers to complex everyday situations. The source separately notes that this question remains open.
editorial commentary
Why it matters
Likely consequence — increased interest in virtual training of controllers as a way to scale robot testing. The next observable signal will be publication of detailed test results in more complex and diverse conditions. Substantial uncertainty is related to the fact that current material does not disclose the evaluation methodology and is based on a single source.