Bear Robotics cuts KR1 robot model latency 8x with NVIDIA tech

Bear Robotics reduced inference time for a robotics foundation model on its KR1 humanoid from 500 ms to 60 ms, roughly 8x, using NVIDIA Isaac OS, Jetson AGX Thor and TensorRT. Its own model improved from 72.3 ms to 17.2 ms. The KINO-2 learning platform was unveiled at NVIDIA GTC Berlin and trains the robot from human demonstrations.

• Isaac GR00T inference latency cut from 500 ms to 60 ms
• Bear's own model improved from 72.3 ms to 17.2 ms
• Uses Jetson AGX Thor, Holoscan SDK 5.0 and quantization
• Bear has deployed over 16,000 robots at 5,000+ locations