Perceptron releases Mk1.5 model for embodied AI agents
The startup says one model flies drones and controls quadrupeds without retraining, and claims its egocentric-vision score beats rivals that cost 25 times as much to run.
Perceptron AI released Mk1.5, a single intelligence layer for embodied agents, the company said. It can fly drones, control quadruped robots, power smart glasses, track objects and reason visually, across devices, without platform-specific retraining, according to Perceptron.
https://x.com/perceptroninc/status/2103508527813669193
A Perceptron researcher posting as AkshatS07 said Mk1.5 is the startup's "first step towards agentic physical world foundation models." It draws on new work in reasoning and reinforcement learning, the researcher said. The model responds 2 to 5 times faster than the previous Mk1 while matching or beating its accuracy, according to the same post.
A second researcher, posting as ArmenAgha, said Mk1.5 posts the company's best score yet on egocentric understanding, the task of reading video from a first-person camera. The model beats rival systems that cost 25 times as much to run, the researcher said. The same account said Mk1.5 combines audio and video training, demonstrated by predicting when a passing truck is closest to the camera using sound alone.
The company said Mk1.5 can also call external tools, such as web and reverse-image search. It can split tasks across sub-agents it dispatches itself, illustrated in a demo identifying allergens in a basket of snacks. None of the performance claims has been independently verified.