Predictive representations
Which representations of video observations help anticipate change? We study feature prediction and task-specific representation evaluation.
KINEJING / WORLD MODELS
KineJing world-model research
At eighteen, we are exploring the physical world through world models.
KineJing (勘境 / 勘境智能) is the brand of 合肥勘境智能科技有限公司. We research predictive representations, action-conditioned dynamics, planning and evaluation. Explore our research questions, public engineering and experiment records.
01 / KINEJING
Young minds. A world to shape.
Curiosity gives us direction. We turn ideas into code, test them through experiments, and let the work speak for itself.
- Brand
- KineJing / 勘境 / 勘境智能
- Legal company
- 合肥勘境智能科技有限公司
- Team
- Built by eighteen-year-olds.
- Public records
- GitHub ↗Hugging Face ↗
02 / WORLD MODELS
Understand change. Explore possible futures.
A world model learns representations and dynamics from observations. Given a state and possible actions, it can predict subsequent change so an agent can compare possible outcomes before acting.
World Models studied compressed spatial and temporal representations for an agent. V-JEPA 2 demonstrates predictive representation learning from video and adaptation with robot data for planning. These are field references, with their authors and results credited separately.
03 / RESEARCH QUESTIONS
From observation to action
Action-conditioned dynamics
How do possible outcomes change when an action changes? We study the relationship between actions and state rollouts.
Planning and evaluation
How can predictions help compare candidate actions under a defined task and budget? We record outcomes, failures and compute conditions together.
04 / PUBLIC RECORDS
Work you can inspect
Project directory ↗Public code and reproduction checks
H-JEPA · KineJing engineering extension
Opt-in candidate retention, action-gradient and rollout-cache changes on upstream H-JEPA, with original authors and licences retained. Paired evaluation tools and initial records are inspectable; the one-task pilot does not show a success-rate gain, and no new trained H-JEPA weights are released.
Inspect the record ↗KineJing-recorded local simulation
Panda dynamics and planning experiments
KineJing studies small state dynamics, sample budgets and CPU planning efficiency in MuJoCo / Franka Panda pushing. Public records include paired conditions, successful and rejected branches, and actual simulator video. Real-robot validation and independent reproduction remain pending.
Inspect the record ↗Six separate evaluation protocols
Public experiments and comparisons
Published charts cover PushT planning objectives, representation readouts, feature prediction and controlled H-JEPA kernel tests. Data, added supervision and evaluation conditions accompany each record; the protocols are interpreted separately.
Inspect the record ↗Public figure and demonstration index
Hugging Face research companion
KineJing Research brings together published figures, demonstrations, reproduction entry points and source notes. This static research presentation is distinct from a trained-model weight release.
Inspect the record ↗05 / DEMONSTRATION SOURCES
Demonstrations and their sources
DROID robot recordings
Original recordings are by the DROID Dataset Team, supplied through NVIDIA Cosmos. KineJing contributes CPU optical-flow diagnostics, synchronization checks and presentation; OpenMDW-1.1 source terms are retained.
Source and scope ↗FourRoom expert-trajectory replay
The expert reference is released by the original H-JEPA authors. KineJing executes environment replay and coordinate checks with CC BY 4.0 attribution; this demonstrates trajectory reproduction rather than a new model-planning evaluation.
Source and scope ↗Panda simulation video
KineJing records actual simulator frames showing a single rollout from the protocol’s preset first seed. Robot assets come from DeepMind Menagerie under Apache-2.0; experiment scope and media checks accompany the record.
Source and scope ↗06 / EVALUATION
Evaluate under explicit conditions
Engineering tests, simulation rollouts and model-quality evaluations answer different questions. Our public records keep outcomes with their scope so the results can be checked.
- Specify the task, scene count, train/test split and whether a test was previously inspected.
- Report shared predictors, sensors and action budgets, alongside added labels or physical priors.
- State hardware, threads, warm-up and measurement scope; separate kernels from end-to-end execution.
- Retain failures, non-improvements and rejected branches with their inspectable records.
07 / LET’S TALK
Create new possibilities. Together.
We welcome world-model research, reproduction and collaboration discussions, and investment enquiries about the team and its public research progress.
Sources reviewed