PROJECT / Open-source adaptation · upstream fork

H-JEPA · KineJing

Optional planning and runtime engineering on H-JEPA, retaining original authors, defaults and reproduction checks.

Current stage: Public engineering extensionGitHub ↗

01 / Overview

H-JEPA · KineJing

Optional planning and runtime engineering on H-JEPA, retaining original authors, defaults and reproduction checks.

02 / Research question

Research question

How can evaluated feasible candidates be retained and unnecessary rollout computation reduced?

03 / Approach

Approach

Optional retain-best candidate selection, action-only gradients and guarded action-embedding caching. Each switch is independent; the default path is retained.

04 / Architecture and scope

Architecture and scope

This page describes research scope. Detailed architecture, interfaces and internal experiments are private.

05 / Current stage

Public engineering extension

06 / Inspect the evidence

Inspect the evidence

View evidence ↗

07 / Limitations

Limitations

An attributed upstream extension. No new trained H-JEPA weights; the initial matched one-task pilot does not show a success-rate gain.

08 / DISCLOSURE

Public information

https://github.com/kineworld/H-JEPA-KineJing ↗

Upstream project ↗

09 / Next steps

Next steps

Verified milestones will be added to the public results pages.