GigaWorld-0
GigaAISoftware / Data / Models
GigaWorld-0
The world's first landmark work proving that world-model-generated data can improve real physical robot performance; 1.5k+ stars on GitHub open-sourced
Low-cost and high-precision, collecting real-scenario multimodal data at large scale and low cost
Developer version, in accelerated delivery, building an ecosystem with the developer community
The world's first landmark work proving that world-model-generated data can improve real physical robot performance; 1.5k+ stars on GitHub open-sourced
Embodied brain (VLA): GigaBrain-0 ranks No.1 globally on RoboChallenge at 51.67% success (nearly 10 percentage points ahead of pi-0.5); GigaBrain-0.5M* is the world's first world-model-based reinforcement learning embodied foundation model
Home general-purpose humanoid robot, first product of the SeeLight sub-brand, carrying the GigaBrain embodied brain; handles meal fetching and prep, tableware tidying, coffee making, and other tasks; won a 100-unit order for deployment at the Optics Valley Zhiyu community in Wuhan
Data collection hardware: low-cost real-robot collection, handheld collection, and ego-centric collection, forming a five-layer data collection system
All-modal real-time world model + agentic model + AI-native creation engine, for film, gaming, and interactive content creation
Autonomous driving world model, the world's first world model for the physical world, serving 30+ leading automakers and autonomous driving companies at home and abroad
Ackermann-architecture chassis, fast with high payload; the HUNTER 2.0 runs at 10 km/h with 150kg payload and automotive-grade independent suspension, suited to outdoor long-distance cruising and scaled-down autonomous driving validation
Tracked chassis with IP67 dust and water protection, handling complex terrain for mine inspection, emergency rescue, and other harsh environments
Four-wheel four-steer chassis, highly flexible, supporting crab steering and high-mobility configurations
Four-wheel four-steer omnidirectional chassis (Ackermann steering / spot turning / crab travel), an excellent choice for mobile manipulation in tight spaces, suited to VLA deployment
Multimodal modular chassis supporting multiple motion mode switches, education kit
Master-slave remote-controlled data collection, ±0.1mm repeatability, for embodied AI data collection; paired with Pika grippers, a desktop ALOHA dual-arm solution, and a PIPER dual-arm kit
First-generation dual-arm home service robot; started paid home rental trials in Sep 2025
Five-finger dexterous hand, 25 DoF, 17 tendons, integrated arm-hand design, whole hand only 85mm long (per earlier notes)
Delivery for restaurants, hotels, retail, and other scenarios; BellaBot Pro integrates an AI foundation model with multimodal interaction, supporting intelligent inspection and restocking and crowd recognition
The T300 shipped over 4,000 units in its first year on the market; the T150 passed 1,000 units within 4 months; announced entry into the unmanned forklift market in May 2026
Full-size wheeled dual-arm embodied robot body, priced from RMB 320,000; executes full-scenario, all-terrain, full-space tasks
Compact wheeled dual-arm version, 60cm arm length, suited to stable single-direction work
Based on the 3D diffusion policy DP3 algorithm; 85% success from 40 demonstrations
All-round manipulation AI robot (flagship); three iterations since launch ('strongest manipulation performance' to 'expert-level operation'); once reposted with comment by Figure AI's CEO and praised by the Physical Intelligence team
Product strategy: the S series for flagship all-round manipulation and the T series for mass adoption, forming a complete product matrix
Core AI technology; a world model built on physical causality as the first principle, with original breakthroughs across three dimensions: physical reasoning, data efficiency, and cross-embodiment transfer
A dual-system architecture that introduces physical-causal modeling on top of mainstream VLA (VLA + world model fusion)
Data system, upgraded from the data loop: structured processing, task explication, and failure-sample accumulation turn raw data into core assets for causal learning
An industry first; uses real physical outcomes to precisely distinguish a model's true generalization ability