LG-controlled Bear Robotics launches KINO-2, cuts GR00T inference on KR1 humanoid 8x

Bear Robotics said in a company announcement on Oct 21, timed to NVIDIA GTC Berlin, that it has launched KINO-2, a platform that collects task demonstrations, trains and refines models, and evaluates the resulting skills on its KR1 humanoid before a customer pilot. KINO-2 supports both Bear's in-house models and external foundation models through shared training, evaluation and deployment tools.

The numbers Bear disclosed

Using Isaac OS with TensorRT and model quantization, Bear said GR00T inference fell from about 500 milliseconds in its first unoptimized deployment to about 60 milliseconds, roughly 8 times faster, in internal measurements; its in-house model went from 72.3 to 17.2 milliseconds, a 4.2 times speed-up. A technical paper on Bear's site gives 56.6 milliseconds median GPU time per action chunk for its optimized GR00T 1.7 build on Jetson AGX Thor, down from 92.2 milliseconds. Bear says it has deployed more than 16,000 robots, mostly service robots, at over 5,000 customer locations. Current manipulation pilots are supermarket returns and warehouse crate handling; at NVIDIA's booth KR1 picks bottles into a crate from a warehouse-system order, delivers it and returns with an empty one. No customer names, prices or KR1 volumes were given.