Core Technology & Innovation
Full-stack Physical-AI Model Architecture
Built on data and simulation, driving industrial cognition and task execution
Drone
Quadruped Robot
Rail Robot
Camera
Industrial Equipment
Multi-agent, multi-device collaboration
Execution constraints → reliability guarantee
Multimodal analysis → task planning
Complex long-horizon task execution
Multimodal data acquisition
Autonomous generation of industrial environments
Task orchestration | Safety constraints | On-site execution
S-VLM / LT-VLA / industry vertical models
Perception | Understanding | Reasoning | Planning
AutoEdge / AutoWorld
Real-world data | Edge-cloud collaboration
Simulated training | Environment generation
Physical AI reshapes next-gen industrial intelligence
-
Perceive the World
-
Understand the World
-
Act on the World
Full-stack Physical AI Architecture
A three-tier physical-AI stack built for industrial demands
JX-Phi Universe — three-tier physical-AI model architecture
-
Tier 1: Data & Simulation Foundation
Problem to solve: "Where does industrial-AI data come from?"
-
Tier 2: Industrial Cognition & Decision Models
Core model: JX-Phi-0.3 Physical-AI Foundation Model (S-VLM + LT-VLA + industry vertical models)
Problem to solve: How does AI "understand" the industrial site and "plan" actions?
-
Tier 3: Deliverable Task System
Core paradigm: Harness · Domain Controller
Problem to solve: How are model capabilities turned into executable on-site "tasks"?
How Physical AI Reliably Delivers Industrial Tasks
System-level solution: closed loop from perception to action, one architecture across diverse high-value industrial scenarios
Power Grid
Power Gen
Chemical
Oil & Gas
Steel
Rail Transit- …More…
- JX Grid Agent
- JX Power Agent
- JX CE Agent
- JX Petro Agent
- 10+ Industry Agents
-
Central BrainCentral Brain
-
One-brain-multi-body engineHeterogeneous terminal onboarding | multi-body scheduling | task allocation | state sync -
Industrial Task HarnessLong-task decomposition | safety rule constraints
JX-Phi (site-level S-VLM 100B) Large vision-language model — multimodal analysis + global task planningDomain Controller
-
-
Controllable Terminals
Supports 100+ device types - Humanoid
- Quadruped
- Drone
- Video Sensor
- Rail-mounted Robot
Embodied Compute Module
-
Jiangxing Data &
Compute Center
Top-tier Academic DNA × 8 Years of Industrial Depth
Academia, industry, patents — a triple technology moat
-
450+
-
60%
-
27
-
1000+
-
8 yrs