Bear Robotics, which has put more than 16,000 mostly serving robots into over 5,000 customer sites, wants its humanoid to handle goods, not just carry them. At NVIDIA GTC Berlin on Wednesday, Oct. 21, 2026, the Silicon Valley company launched KINO-2, a robot-learning platform that turns task demonstrations into deployable manipulation skills on KR1, the humanoid it gained when it bought Kinisi Robotics in 2026.

What it does

  • Bear collects demonstrations from workers, trains and refines models, and tests the skills on KR1 before customer pilots.
  • Current manipulation pilots include supermarket returns and warehouse crate handling.
  • At the NVIDIA booth, KR1 takes a warehouse-management-system order, picks bottles into a crate, delivers it to a drop-off point and returns with an empty crate.
  • The platform runs Bear's own models and outside foundation models, including NVIDIA Isaac GR00T.

The speed claim

KR1 runs NVIDIA's Isaac OS runtime (with Holoscan SDK 5.0) on a Jetson AGX Thor computer. Using Isaac OS, TensorRT and model quantization, Bear says it cut GR00T inference time from about 500 milliseconds in its first unoptimized deployment to about 60 milliseconds, roughly 8 times faster, and its in-house model from 72.3 to 17.2 milliseconds (4.2 times). These are Bear's internal measurements, as the Korea Economic Daily also reported. NVIDIA used the same event to release the first public beta of Isaac OS.

"Demonstrations let us capture that knowledge and teach the robot," said Bren Pierce, Bear's chief robotics officer. Bear did not give a price or a date for wider commercial rollout.

Why it matters for Tesla watchers

Learning from human demonstrations is also how Tesla trains Optimus. See Musk on Optimus hands and AI5 and Schaeffler and NVIDIA's humanoid actuator work.