Does Jev output joint angles directly?

No. Jev receives compact text and JSON facts and returns intent plus positive, negative, or stay per axis and open, close, or stay for fingers, each with probabilities and confidence. The harness maps labels to Cartesian targets, DLS IK, filtered joints, and physics.

What machine and key setup is needed?

Python 3.12 is the tested baseline, with an EGL-capable GPU recommended for the live panel and 1080p60 export. Set TYPESAFE_API_KEY or JEV_API_KEY, or pass a key or key file to start.sh, then open the paused panel port.

How can I audit a reference episode?

Each run saves replay video, joint-state recording, both Jev responses with before and after observations, task and stability report, and export metadata. The first API connection warms before recording so cold start does not inflate task time.

What task does the harness demonstrate?

The demo moves an orange block over a blue barrier to a tan pad before withdrawing. Intent options are approach, grasp, lift, carry, lower, release, withdraw, and finish; no Jev weights download to the simulation machine.

How many TypeSafe calls does each policy update make, and what do they decide?

Two sequential calls. The first chooses the manipulation intent; after the harness reads MuJoCo again, the second chooses the axis and finger moves from the selected intent and fresh contact facts.

What choices make up the intent question?

approach, grasp, lift, carry, lower, release, withdraw, and finish. Each Choice returns the selected label, probabilities, and a confidence value.

What license covers the harness versus the robot assets?

The launcher and harness are released under the repository license, while the included Franka Panda assets retain their upstream license in assets/franka_emika_panda/LICENSE.

CommunityJev Demos & Experiments

Jev-as-Policy

Self-contained MuJoCo reproduction that uses two sequential Jev Choice calls per update for manipulation intent and axis plus finger moves, with local IK and servo doing execution.

See highlights

Open github.com

Stars: 48
Language: Python
License: MIT
Published by: Community
Author: YuanKJing
Last verified: Oct 2026

YuanKJing/Jev-as-Policy

User repository on GitHub

View on GitHub

The highly anticipated open-source repository for JEV as Policy enables one-click setup of the simulation environment. Evaluations of Astra + JEV on benchmarks such as RoboTwin will also be released soon.

Forks: 2
Last pushed: Updated Sep 2026

Repo stats from the GitHub API, cached Oct 2026.

About this project

Text state → Jev intent → Jev motor choices → local IK → MuJoCo. This repository is a self-contained reproduction of the public Jev as Policy control structure. It includes the MuJoCo scene, Franka Panda assets, TypeSafe/Jev integration, continuous Cartesian servo, recording pipeline, and the studio-style web panel. The included panel recording is a reference result from this exact harness: Watch the 1080p / 60 fps panel demo.

Quoted from the project's README. Read the full README.

Tagscommunity jevbest catalog-import

Works forJev for Classification Jev for Evaluation

Used inJev for Software & DevTools

Languages: Python

Highlights

Sends MuJoCo geometry and contact facts as text and JSON, never images, meshes, or joint trajectories.
Makes two dependent TypeSafe calls per update: one eight-way intent Choice, then four axis and finger Choices.
Executes with continuous Cartesian servo, DLS IK, filtered joint targets, and contact-aware physics locally.
Records decisions, observations, joint states, stability reports, and panel video under timestamped experiment folders.

Watch out

Remote TypeSafe decisions with local physics ownership; video is rendered from recorded states on the wall-clock timeline, not edited to hide API waits. Copy reviewed against the repository README on 2026-10-04, not from running it.

Questions about this entry

Short answers drawn from the linked project or page.

More Jev resources like this

openthai-systemone

Open Thai plus English decision model that answers typed choice, score, and noul questions about a state in one forward pass.