User repository on GitHub
Camera-only autonomous drone in MuJoCo with a small judgment model (TypeSafe Jev) in the loop at 2.5Hz
- Forks
- 40
- Last pushed
- Updated Sep 2026
Repo stats from the GitHub API, cached Oct 2026.Homepage
Jev directory
CommunityJev Cookbooks
Jev Drone
A camera-only MuJoCo quadrotor that uses Jev for tactical maneuvers at 2.5 Hz while classical perception and a 50 Hz reflex layer keep safety outside Jev, clearing a five-station course.
- Stars
- 277
- Language
- Python
- License
- MIT
- Reading time
- 3 min
- Published by
- Community
- Author
- RomanSlack
- Last verified
- Sep 2026
About this project
An autonomous quadrotor flies a five-station obstacle course in MuJoCo using only its onboard camera. A small judgment model (TypeSafe's Jev) sits at 2.5 Hz and decides what the situation means; everything time-critical stays in ordinary code. The drone is the Skydio X2 from MuJoCo Menagerie. a real airframe with four rotors, flown by a geometric controller. Nothing about the flight is scripted.
Quoted from the project's README. Read the full README.
Highlights
- Jev answers maneuver (Choice), risk (Score), and target_truly_lost (Noul) in one call.
- A typical 65-second flight uses about 110 calls at 0.11 s median latency.
- The baseline without Jev reaches 17.7 m and gets stuck at station 2; with Jev it clears 77.5 m.
- A 500 Hz controller, 50 Hz safety reflex, and 15 Hz vision loop all run without Jev.
Watch out
Results come from a single 65-second run with large variance, and a matched 3-seed test in an earlier arena showed no advantage.
Questions about this entry
Short answers drawn from the linked project or page.
No, and it cannot. Awareness is numpy over depth and segmentation buffers at 15 Hz with zero Jev involved; classical vision turns those into range sectors, the blocking height, and target bearing. Jev takes JSON, not pixels. Flight control runs at 500 Hz and the safety reflex at 50 Hz, both in code, and code can veto Jev.
One 65-second run, not a seed-matched average, and the README says so. On an earlier, simpler arena a matched 3-seed comparison showed no advantage at all for Jev, because the extra maneuvering cost target visibility. The claim it supports is narrow: the greedy baseline is structurally incapable of flying over an obstacle and Jev supplies that maneuver.
It runs at roughly 2.5-3 Hz and is advisory only. Code decides when to ask at all, so an open corridor with the target in view costs nothing, and unchanged scenes are fingerprinted to reuse the last judgment. A 50 Hz reflex overrides Jev: it can propose climb, but code refuses unless the measured top edge is inside the aircraft's climb ceiling.
Because the state could not support the answer. Five horizontal range sectors carried no vertical information, so 'fly over it' was not an inferable option. Adding the obstruction's top-edge height, whether that edge is visible to the camera, and the aircraft's own climb ceiling moved climb from never-chosen to p=0.93. That was a state-design bug, not a model failure.
Run `./setup.sh` to build the virtualenv and fetch the Skydio X2 model, copy `.env.example` to `.env` and put your key in it, then `.venv/bin/python run.py --seconds 65 --seeds 1`. On a headless box set MUJOCO_GL=egl instead of glfw.
Yes. The ablation runs with `--no-jev --fast`, and `replay.py` re-renders a recorded `course.mp4.tape.npy` file without flying again or calling the API. A video run writes that tape file specifically so you can re-cut visuals later for free.
Jev sits at roughly 2.5 Hz and answers three questions in one call: maneuver as a Choice, risk as a Score, and target_truly_lost as a Noul. Code decides when to ask, unchanged scenes reuse the last judgment via fingerprints, and a typical 65-second flight costs about 110 calls.
Reactions & coverage
Posts, threads, and videos about this entry from around the web.
Hacker News: Pick Verify: a single decision at the warehouse pick station catches confusable picks.
Show HN: Verifying robot picks with a decision modelRead the thread on Hacker NewsX: A dual-arm robot with Jev in the middle layer
YouTube: Dan Makes Things drives an ESP32 robot with Jev decisions.
More Jev resources like this
An experimental controller that has Jev choose NES inputs to play Super Mario Bros. from structured emulator telemetry instead of screenshots, logging every decision with probabilities and latency.
Jev Cookbooks#community#python#games
Community
fhshaik/typesafe-marioPython
fhshaik/typesafe-marioPythonA fine-grained robot control study loop for LIBERO tasks: Jev layers an intent, a contact or motion family, and one of 27 control inputs, while reversible MuJoCo previews evaluate each candidate before execution.
Jev Cookbooks#community#python#robotics
Community
Dimweaker/jev-liberoPython
Dimweaker/jev-liberoPythonA MuJoCo and Franka Panda workbench where Jev chooses the next manipulation step for pick-and-place, stacking, and obstacle-carry tasks, with rule baselines and side-by-side model comparison.
Jev Cookbooks#community#python#robotics
Community
FBddcz/embodied-jevPython
FBddcz/embodied-jevPythonA MuJoCo city simulation where one, two or four camera-observed drones pursue a target: each role asks Jev role-specific typed questions while deterministic mission navigation, safety guards and motor control run independently at 50 Hz.
Jev Demos & Experiments#community#github#python
Community
YongLD/Multi-UAV-JEVPython
YongLD/Multi-UAV-JEVPythonTwo-stage Jev control of a Franka Panda in MuJoCo: Jev selects a task intent, then X/Y/Z directions and a gripper command executed against real contacts, with 43 of 50 Jev episodes succeeding versus 48 of 50 for a rule baseline.
Jev Cookbooks#community#python#robotics
Community
lykycy123/RoboJEVPython
lykycy123/RoboJEVPythonA community curriculum of Jupyter notebook recipes that teach typed decisions with Jev, TypeSafe's System One model. Each recipe takes one bounded Choice, Noul or Score decision over text or image state, spanning first classification through closed-loop control.
Jev Cookbooks#community#github#python
From the Jev guides
Original write-ups on what this entry is for.
From the Jev AI model community
Posts from builders shipping with Jev right now.