Why a 7‑Month‑Old Startup Claims Human‑Like Robots Are Key to General Embodied Intelligence

The launch event featured KAI, a humanoid robot that introduced itself and hosted its own product announcement, showcasing that a robot can independently present its capabilities.

The author argues that high anthropomorphism is essential because the world is built for humans; a robot with human‑like dimensions can directly leverage massive human motion data, as noted by MIT researcher Rodney Brooks and other industry viewpoints.

Physical specifications : KAI stands 173 cm tall, weighs 70 kg, and has a head‑to‑body ratio of 1:8.5. It offers 115 full‑body degrees of freedom—the highest publicly reported for a humanoid—and a hand with 36 DOF (22 active, 14 passive) featuring a single‑direction self‑locking mechanism. Its tactile skin covers 80% of the body with 18,000 touch points, detecting forces as low as 0.1 N. Power is supplied by a custom semi‑solid battery providing over 3 hours of continuous operation.

Design philosophy : Engineers start from human biomechanics, reverse‑engineering joint mechanics and extending the approach to body proportions, physique, fitness, and sensing, ensuring the robot’s dimensions align with human environments and reduce the uncanny valley effect.

Cognitive architecture : KAI employs a three‑model loop—world model predicts the outcome of actions, action model generates candidate motions, and evaluation model scores candidates on task proximity, safety, and feasibility—embodying a “plan‑then‑execute” decision process.

Data collection : The company has amassed 100,000 hours of first‑person video and pose data using the KAI Halo head‑mounted device, crowdsourced across more than 20 real‑world scenes. An automated infra platform cleans, remaps, and balances the data before feeding it to the training pipeline.

Training strategy : A three‑stage regimen is used. Pre‑training leverages internet, simulation, and self‑collected data to build general physics knowledge. Bridge training adds contact‑force details using UMI datasets and glove recordings. Post‑training incorporates real‑world tele‑operation data to align the model with the physical robot.

The article concludes that while KAI demonstrates a tightly integrated hardware‑software system where high‑fidelity embodiment fuels a powerful world model, the approach faces significant hurdles: the engineering complexity of 115 DOF, sophisticated tactile hands, and custom power solutions translates to high cost and a long path from lab showcase to widespread, reliable deployment.