Bear Robotics introduces KINO-2 for robot learning

From demonstrations to manipulation

Bear Robotics announced KINO-2 on October 21, 2026 as a platform that turns task demonstrations into deployable manipulation skills for KR1, Bear's humanoid robot. The launch announcement describes an extension of the company's work from transporting goods to handling objects within a customer's process.

Bear collects demonstrations, trains and refines models, and evaluates the resulting skills on the robot before deployment in a customer pilot. The company says active pilots include supermarket returns and warehouse crate handling.

Running learned skills onboard

KINO-2 provides shared training, evaluation and deployment tools for Bear's in-house models and external foundation models. Bear says this lets different model architectures reuse its existing navigation, fleet management and workflow systems, while model selection depends on the task and onboard computing constraints.

The platform was developed through engineering collaboration with NVIDIA. According to Bear's technical post, NVIDIA Isaac OS underpins its perception runtime, connecting camera processing and model inference on NVIDIA Jetson AGX Thor. An accelerated pipeline and NVIDIA TensorRT help run more capable models onboard.

Kinisi, Bear's wheeled humanoid

Bear separately lists Kinisi as a wheeled humanoid mobile manipulator. Its product page states a speed of up to 14.4 km/h and an IP65 rating, while payload and reach remain to be confirmed. Bear has not said whether Kinisi is the KR1 named in the KINO-2 announcement.

Specs as published are on RoboHub's Kinisi page. See also the Bear Robotics brand page, Servi and the humanoid robots directory.