World Models
Model how environments change over time and in response to actions.
勘境 KINEWORLD
Explore change. Envision what comes next.
Use world models to study change and explore possible outcomes before acting.
KineWorld researches action-conditioned world models, connecting representation, prediction, planning and evaluation. Public projects show code, progress and limitations.
WORLD MODELS FOR PHYSICAL INTELLIGENCESCROLL TO ENTER ↓
01 / THE RESEARCH QUESTION
If we act, how might the world change?
KineWorld researches action-conditioned world models: learning representations from observations, anticipating the effects of actions, and exploring how predictions can inform planning.
This is a research path. Each capability needs validation against specific tasks, data and baselines.
Explore research ↗02 / THE NAME
Survey the unknown. Understand the world.
The name KineWorld begins with a way of working and the world it studies.
勘Survey, measure, verify. First, understand what is here.
境Environment, boundaries and possible futures. Then, understand how they change.
About KineWorld ↗03 / What is a world model?
From seeing to understanding change.
Perception asks, “What is here now?” A world model asks, “What might happen if I take this action?”
Prediction is never certainty. Research must record assumptions, failures and boundaries.
04 / Research path illustration · RESEARCH DIRECTION
From observation to feedback.
- 01Observation
- 02Representation
- 03Action-conditioned prediction
- 04Future rollout
- 05Planning
- 06Action
- 07Feedback
This is a research path, not a list of mature capabilities.
05 / RESEARCH PILLARS
Research pillars
Representation Learning
Extract states from observations that support prediction and action.
Action-conditioned Prediction
Study how possible outcomes depend on the chosen action.
Spatial Understanding
Build useful representations of position, structure and relationships.
Causality & Dynamics
Distinguish co-occurrence, change and consequences while keeping claims testable.
Planning & Control
Explore how predictions can inform goal-directed action selection.
06 / PROJECTS
Selected projects
First-party project ↗
KineJing
Evidence-bound experiments for action and world-model research, including a lightweight motion baseline, a visual feature path, and a small trained three-view action-feature predictor.
Research previewFirst-party project ↗kine-jepa
Research code exploring compact action-conditioned latent world models for planning and control.
Research codeFirst-party project ↗kine-bench
Evaluation and evidence gating for claims about world-model planning.
Research tooling07 / OPEN RESEARCH
Evidence before claims
We aim to connect research claims to inspectable code, experiments, evaluations and limitations.
GitHub ↗08 / WHAT COMES NEXT
Think through the next step.
From reproducible research toward systems that can compare possible futures before acting.
Build inspectable foundations
Public projects cover action prediction, latent dynamics and evaluation. The repositories and model cards define their current status.
Test longer horizons
We plan to explore shared data and evaluation interfaces, action-conditioned multi-step prediction and planning experiments.
Anticipate before acting
Test in controlled settings whether models can help systems understand consequences, choose actions and account for uncertainty.
The latter two are research directions, not delivered capabilities or timeline commitments; the path will change with evidence.
09 / UPDATES
Recent updates
KINEWORLD / 勘境