Skip to content
Metatext Decide — typed decisions for your agents, with a probability on every answerGet early access
Metatext
Sign Up

jev-drive

Research prototype testing Jev as a semantic supervisor that tells a driving policy to continue, observe, replan or defer.

ServiceresearchMIT
View on GitHub ↗

Overview

JevDrive tests where Jev can help an existing autonomous driving policy choose to continue, observe, replan or defer. A local controller and feasibility checks own vehicle execution; a separate vision model turns camera images into text, and Jev receives structured text plus numerical facts. Observation age and candidate checks run before any recommendation reaches the existing policy.

It is a working research prototype: the live Jev client, recorded-image comparison, response gate and evaluation tools exist, and a 27-image GPU diagnostic has been run, but closed-loop CARLA or BeamNG integration and any driving improvement are not demonstrated. In that diagnostic, the README reports Qwen2.5-VL scoring directly at a median 275 ms per image versus about 4 seconds for a caption-then-Jev pipeline, with seven Jev failures out of 54 attempts. The author concludes Jev suits occasional recovery decisions, not per-frame control.

What the decision model decides

TypeQuestionAnswers
ChoiceShould the vehicle continue, observe, replan or defer?continue · observe · replan · defer

How it works

  1. Camera images become vision observations; state provides numeric facts
  2. Semantic supervisor asks Jev for a recovery decision
  3. Check observation age and candidate feasibility
  4. Pass the decision to the existing driving policy

Strengths

  • Keeps execution and feasibility in deterministic code
  • Honest measurement with stated limits
  • Clear guidance on where Jev fits

Caveats

  • No closed-loop driving results
  • Caption-to-cloud pipeline too slow for per-frame use
  • Some Jev calls failed with 503s or timeouts in testing

At a glance

Type
Service
Status
research
Language
Python
Runtime
Research prototype with GPU vision model
Has code
yes
License
MIT

Use cases and integrations

How curated lists describe it

  • Jev + autonomous driving: structured decisions, multimodal baselines, recovery research, and measured API diagnostics.

    logicrw/awesome-jev-projects

Frequently asked questions

What does Jev decide?

Whether the driving policy should continue, observe, replan or defer.

Does it improve driving?

Not demonstrated; closed-loop simulator tests have not been run yet.

Is it fast enough for real time?

Not per frame; the author targets occasional recovery decisions.

Related projects

  • OmniJev

    Multimodal robot decision SDK: camera images plus text go to Qwen3.5-9B with a corrective head that picks the next preset arm skill.

    BenchmarkresearchPython
  • amigos-jev

    Free hosted Jev-style multimodal decision model API (about 20B active) that picks actions from images and options, aimed at robot control.

    Servicebeta
  • jev-spatial

    Independent open model on Molmo2-ER that answers spatial questions by choosing from fixed options with one shared classifier head.

    ProjectresearchPython
  • OmniJev

    OmniJev: omni-modal 4B decision model answering typed questions about images, video, screens and robot scenes in one pass.

    ModelresearchPython
  • jev-askable-arm

    Simulated Franka arm that follows plain-English goals by having Jev chain about 30 hardcoded motion primitives, one per step.

    Agent skillprototypePython
  • jev-drone

    Camera-only MuJoCo quadrotor where Jev makes 2.5 Hz tactical judgments while code keeps flight control and a safety veto.

    ProjectprototypePython