Can Superdimensional Power’s Full‑Stack Embodied AI Turn Robots into Users of Cloud‑Based Large Models?

At WAIC, Superdimensional Power showcased a full‑size humanoid robot playing ping‑pong and a dual‑arm robot opening a beverage can with zero errors, demonstrating the need for rapid visual perception, trajectory prediction, motion planning, and whole‑body control.

The robot’s skill acquisition relies on a data chain where roughly 90% of the experience comes from first‑person human recordings. The model first learns task flow from human data, converts human viewpoints and motions into robot‑usable representations, and finally refines fine‑grained actions and failure correction with real‑robot data.

Human and robot perception differ: head‑mounted cameras move with gaze, while robot cameras have fixed positions; human shoulders, waist, and wrists cooperate naturally, whereas robot joint topology is fixed by the hardware. Consequently, understanding a task does not guarantee accurate execution without bridging viewpoint, body structure, action space, and physical feedback gaps.

Superdimensional Power’s product line—KAI Bot, KAI Hand, KAI Halo, world model, and AI infrastructure—aims to close this loop. KAI Bot stands 173 cm tall, weighs ~70 kg, and offers 117 DOF with ~80% tactile skin coverage, emphasizing added freedom in the neck, shoulders, and waist to match human‑scale environments and capture full‑body human motion.

“We are a company of embodied intelligent large models.” – Luo Ping, co‑founder and professor at HKU.

High DOF expands the robot’s reachable workspace and enables direct use of existing tables, shelves, and tools, but it also raises hardware cost, calibration difficulty, fault points, and algorithmic complexity; more human‑like motion does not automatically mean higher task efficiency.

Superdimensional Power incrementally unlocks joints: early algorithms lock new DOF, then gradually enable neck, shoulder, and waist capabilities to manage the control burden of 115 DOF.

Two tactile demonstrations at WAIC highlighted structural robustness of KAI Hand under impact and whole‑body skin sensing of KAI Bot, underscoring safety concerns—robots must avoid accidental harm while eventually needing safe physical contact for assistance tasks.

Data collection involves hundreds of operators recording first‑person video, head motion, hand actions, and object interactions in homes, stores, and factories using devices like KAI Halo Lite. Raw data quickly reaches petabyte scale, but is noisy due to camera shake, occlusions, and varied human habits, requiring extensive pipelines for quality control, 3D reconstruction, pose recovery, and motion retargeting.

“First‑person human data is extremely noisy; we must extract effective samples via algorithms and automated pipelines.” – Luo Ping

The training pipeline is split into three stages: pre‑training on large‑scale human first‑person data to learn scenes and tasks; bridge training that maps human hand and body motions to robot‑centric action representations; and post‑training that incorporates real‑robot data and reinforcement learning to adapt to each robot’s dynamics and recover from failures.

The can‑opening demo validates this chain: the model learns the procedure from human data, transfers it to a dual‑arm robot, and corrects misalignments by detecting deviations and re‑adjusting, proving that human experience can be transformed into robot actions after data processing, bridging, and real‑robot learning.

However, a single demo does not prove scalability; key metrics include success‑rate improvements with and without human data, generalization across different cans and tools, data required for new tasks, and iteration speed after failure feedback.

Superdimensional Power also pursues a world‑model that extends data generation beyond real environments. The system follows three phases—Design the World, Play with the World, and Learn about the World—comprising world, action, and evaluation models. The world model must generate physically consistent data, not just photorealistic video, to be useful for robot training.

Evaluation of the world model hinges on whether generated data improves real‑robot task success and whether virtual failures correspond to real‑world issues.

The full‑stack approach connects hardware, data, models, and deployment in a closed loop: human and robot generate data, the AI infrastructure processes it, models learn capabilities, robots execute tasks, and execution results feed back into training. The three efficiency metrics—data acquisition efficiency, conversion of data to capability efficiency, and cross‑robot capability transfer efficiency—determine the speed of robot evolution.

Luo Ping emphasizes that the goal is not to define a robot’s boundary but to establish a method for continuous robot evolution, making the KAI Bot merely the first “user” of a platform that could serve diverse robot forms.