Armen Aghajanyan and Akshat Shrivastava spent years at Meta’s FAIR lab. Their startup Perceptron AI, founded November 2024 and backed by $21M from Bessemer, just released Isaac 0.5: a 36B dynamic-MoE embodied foundation model that packs video understanding, embodied reasoning, and robot control into one sparse backbone. Weights, inference, and training code — all open.
The numbers that flipped the race
Closed labs owned robotics until now. Isaac 0.5 hits 97.2% average success on LIBERO, edging out NVIDIA GR00T N1.7 (97.0%) and Physical Intelligence’s π0.5 (96.9%). The wilder stat: train once on a single expert demonstration and error rates drop 7-10.5x, versus π0.5’s 2.3-3.1x. Behind it sits 3 trillion multimodal tokens, 1 million hours of video, and 100K hours of robot experience across 35+ systems. Scaling video pretraining cut the teleop data needed for a target from ~5,900 hours to 28.
How you actually use it
Download the weights and run Isaac as a direct control policy on your robot, or feed its visual outputs — object tracking, task-state estimates — into an existing planner. Perceptron is aiming it at manufacturing, logistics, and warehouse floors.
You Might Also Like
- Ricursive Intelligence Just Raised 335m in 4 Months and i Think Nvidia Should be Paying Attention
- Sharpa Robotics North Sharpawave Hand 22 dof 1000 Tactile Pixels per Fingertip 54 Training Boost With Nvidia Gr00t
- Physical Intelligence π0 7 Controls Ur5e Franka and pi Droid From one Cloud api
- Lg Cloid x Nvidia Isaac Seoul Deal Puts a 7 dof Home Robot Inside Isaac sim
- Tutor Intelligence Data Factory 1 100 Robots Generate 10000 Hours of Training Data per Week
Top AI Product 
Leave a comment