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

SafetyJev

Research system that uses a safety-tuned Jev-style predictor to forecast constraint violations of a robot VLA policy and gate actions.

Projectresearch
View on GitHub ↗

Overview

SafetyJev is a research project on making agentic robot manipulation safer with an in-the-loop, safety-centric fine-tuned Jev-style model. It runs the ManiGuard pi0.5 vision-language-action policy, forecasts constraint violations before action chunks execute, and compares predictions with ManiGuard's physics-grounded LTL monitor. A guarded mode checks every constraint before executing a chunk and asks the VLA to regenerate on rejection, with an optional OpenRouter planner that rewrites the instruction using rejection feedback, cameras and bounded history.

The native predictor loop uses a pinned checkpoint trained on unsafe examples, accepts candidates predicted safe on all constraints, and after budget exhaustion executes the candidate satisfying the most trained queries. Connectors include an Open-Jev HTTP Noul connector for a proprio-only ablation and a trained 27B visual SafetyJev using two cameras. The README is candid about status: a single-task smoke test exercised repairs and fallback but checkpoint replay parity did not pass, and current runs are limited to 200 base cases in Isaac Sim.

What the decision model decides

TypeQuestionAnswers
Yes / noWill executing this action chunk violate constraint X within the horizon?—

How it works

  1. VLA proposes an action chunk from the observation
  2. SafetyJev predicts violations for every constraint
  3. Reject and regenerate (optionally with a repaired instruction) if any violation is predicted
  4. Execute accepted chunks; compare with the physics monitor

Strengths

  • Predicts violations before they happen
  • Measured against a physics-grounded oracle
  • Calibration and timing diagnostics included

Caveats

  • Research code; replay parity did not pass in the smoke test
  • Requires Isaac Sim, ManiGuard and GPUs
  • No licence stated

Getting started

python3 -m safetyjev.cli verify-integration --maniguard-root /path/to/ManiGuard

At a glance

Type
Project
Status
research
Language
Python
Runtime
Isaac Sim simulation + CLI
Has code
yes
Has tests
yes

Use cases and integrations

How curated lists describe it

Frequently asked questions

What does SafetyJev decide?

Whether a proposed robot action chunk will violate each safety constraint before it executes.

Is it TypeSafe's Jev?

It uses fine-tuned Jev-style models (Open-Jev and a visual SafetyJev), not the hosted API.

Is it ready for real robots?

No; it is simulation research with open issues noted in the README.

Related projects

  • 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
  • jev-reflex-autonomy-lab

    Multi-drone simulation where Jev makes fast reflex decisions per drone and an optional slower LLM may advise but never takes control.

    AppprototypeTypeScript
  • robo-harness

    SO-101 robot-arm workbench where a Jev decision runner picks bounded joint steps from typed candidates under a spend cap.

    ProjectprototypeTypeScript
  • DriveJev

    Open 4B System I driving model that scores behaviours like cruise, stop, yield and brake from three camera frames each decision.

    ProjectresearchPython
  • JEV Control Your Roboarm

    Experimental robot loop where Jev chooses the next verified action from perception-generated options for a language goal.

    ProjectresearchPython