Robots are getting smarter all the time thanks to all the new models that are being developed for them. Perceptron AI’s Isaac 0.5 is a 36B dynamic-MoE, open-weight embodied foundation model. It relies on video understanding, reasoning and robot control to enable robots to take on challenging tasks. This model is “trained across 35+ robot systems, 100K hours of robot experience and 1M hours of video.”
Perceptron AI unveiled Isaac 0.5, a 36B dynamic-MoE, open-weight embodied foundation model.
It combines video understanding, embodied reasoning and robot control in one sparse backbone, trained across 35+ robot systems, 100K hours of robot experience and 1M hours of video.… https://t.co/boL4WYlcxW pic.twitter.com/iA6jDP0Jdv
— RoboHub🤖 (@XRoboHub) August 28, 2026
As the researchers explain,
Isaac 0.5 also leads all five benchmark families it evaluates, at up to 8.5× lower inference cost than the strongest open comparator.
[HT]