How Data and Algorithms Enable Embodied Intelligence Scaling – GigaAI’s Dual‑Pyramid Physical AGI

From Scaling Bottlenecks

The embodied‑intelligence field has struggled to satisfy the Scaling Law because single data sources cannot simultaneously provide scale, information density, and realism, and current VLA‑style models cannot efficiently encode 3D physics and continuous actions.

Five‑Layer Data Pyramid

The data hierarchy from bottom to top consists of Internet video data, real‑person data, world‑model simulators, synthetic simulation data, and real‑machine data . Each layer addresses a different trade‑off between scale and fidelity.

GigaAI supplies engineering products for every layer: real‑machine data are collected by the household arm robot SeeLight S1 and the low‑cost Maker M01 sensor; synthetic and world‑model data are generated by the self‑built platform GigaWorld‑0; real‑person data are captured at scale with handheld U‑01 and ego‑camera E‑01 devices; Internet video data reuse public sources such as YouTube and the Panda‑70M dataset.

Three‑Layer Algorithm Pyramid

The algorithm hierarchy from bottom to top includes world simulation, action alignment, and experience reinforcement , each mapped to concrete models.

World‑simulation is represented by GigaWorld‑1 , which achieved a composite score of 62.34 on the WorldArena benchmark, surpassing Wan, CogVideoX, Veo 3.1, and Cosmos‑Predict to become the first embodied world model above 60 points.

Action‑alignment comprises the GigaBrain‑0 series and GigaWorld‑Policy . GigaBrain‑0 topped the global RoboChallenge with a 51.67% task‑success rate, beating the π0.5 baseline by nearly 10 percentage points, while GigaWorld‑Policy led the RoboCasa365 ranking, defeating NVIDIA GR00T N1.5 and π0.5.

Experience reinforcement is embodied in GigaBrain‑0.5M , which combines the world model with reinforcement learning to enable self‑evolution of the embodied base model.

Home Deployment as the Next Testbed

Having validated the technical stack, GigaAI turns to real homes. The company launched the sub‑brand SeeLight and introduced the first general‑purpose household humanoid robot, SeeLight S1, claiming it as the world’s first robot operating in actual residential settings.

Hundred‑Robot Deployment

GigaAI secured an order for 100 SeeLight S1 units to be deployed in the Wuhan Optics Valley residential community, marking the first publicly disclosed large‑scale household robot rollout in China and providing a source of long‑term real‑world usage data.

S2: Systematic Redesign

The upcoming second‑generation robot, SeeLight S2 (Q3 2026), will shrink chassis volume by 60%, extend battery life by 70% with hot‑swap capability, and increase operational reach by 40%, aiming to improve usability for continuous service in homes.

12‑Month Roadmap: Three Foundational Models

GigaAI announced a 12‑month roadmap for three successive foundational models: GigaBrain‑1 (Q3 2026) will be the first physical‑AGI model built on the dual‑pyramid framework; GigaBrain‑2 will follow, and GigaBrain‑3 will be trained on 10 million hours of video plus 1 million hours of world‑action data, targeting the “GPT‑3 moment” for physical AGI.

Key Follow‑Up Items

Three aspects merit close monitoring: whether the dual‑pyramid system truly achieves Scaling Law performance, whether the hundred‑robot household deployment yields a closed‑loop data pipeline, and whether the projected GPT‑3‑like breakthrough materializes within the announced timeframe.