Hetu / 河图

Predictive World Model

Status: Preview / Research Roadmap

Hetu is the predictive world-model layer of XUANSHU AI. Its long-term purpose is to predict how a world may evolve from its current state under a candidate action.

Formally, the core problem can be expressed as:

z_(t+1) = f(z_t, a_t)

where z_t is the world state at time t, a_t is a candidate action, and z_(t+1) is the predicted future state.

Why a World Model

Perception describes the present, but an intelligent system also needs to reason about consequences before it acts. A world model provides a basis for:

• action-conditioned future-state prediction;
• multi-step rollout across a prediction horizon;
• uncertainty estimation;
• collision, contact, and risk prediction;
• comparison of possible outcomes before execution.

Hetu is intended to provide these predictive capabilities without taking ownership of perception or robot execution.

Position in XUANSHU AI

Jingwei / 经纬
Perceive the World
        ↓
Hetu / 河图
Predict the World
        ↓
Gongshu / 公输
Act in the World

In one sentence:

Jingwei answers “what is here now,” Hetu answers “what happens next,” and Gongshu answers “how to act.”

This diagram describes the intended product architecture. Jingwei, Hetu, and Gongshu are not yet integrated into a real end-to-end system.

Product Boundary

Hetu is responsible for:

• world-state representation;
• temporal dynamics;
• action-conditioned prediction;
• future-state prediction;
• multi-step rollout;
• uncertainty estimation;
• risk prediction.

Hetu does not provide object detection, annotation, or camera perception; those concerns belong to Jingwei. Hetu also does not provide inverse kinematics, robot control, grasp execution, or manipulation execution; those concerns belong to Gongshu.

Current Status

Hetu is currently a Preview / Research Roadmap repository. This release defines the product concept, research questions, system boundaries, and a possible development path.

There is currently:

• no implemented world-model algorithm;
• no trained model or checkpoint;
• no training or inference pipeline;
• no dataset or data API;
• no production API;
• no claimed benchmark result;
• no end-to-end integration with Jingwei or Gongshu.

The documents in this repository describe intended research and engineering directions, not available model capabilities.

First Technical Target: Hetu v0.1

The proposed first technical milestone is a Tabletop Predictive World Model. In a simple tabletop robotics setting, it would take:

• Current Scene State
• Robot Action

and predict:

• future object position;
• future object pose;
• contact;
• collision;
• future scene state;
• future robot state;
• multi-step future state.

The initial scope should favor structured, measurable state transitions before progressing to learned latent representations. This is a roadmap target only; it is not implemented in the current repository.

Research Direction

The planned progression is:

1. define a structured tabletop state and action schema;
2. establish state-transition prediction tasks and evaluation protocols;
3. investigate learned dynamics, including latent, transformer-based, and object-centric models;
4. extend prediction to multi-step rollout, calibrated uncertainty, and risk;
5. evaluate future interoperability with Jingwei perception outputs and Gongshu action candidates.

See PRODUCT_CONCEPT.md, ROADMAP.md, docs/architecture.md, and docs/world-model.md for the current product and research specification.

Release Principle

Hetu will not claim capabilities that do not exist. Code, models, datasets, metrics, and integrations should be added only when they are real, testable, and documented.