Question 1

Has the Jev stepper actually driven a physical robot, or is it simulator-only so far?

Accepted Answer

Not on hardware. The decision model is off unless you name one with --decision-llm, and the Jev stepper has not been measured on a physical robot. Of seven supported bodies, one runs on real hardware: an SO-101 follower arm over LeRobot, piloted by OpenAI's gpt-6-astra over twelve --goal runs.

Question 2

What does the decision model do that the LLM pilot cannot?

Accepted Answer

It scores a fixed set of options against a state and returns which one with a probability for each, generating no text at all. So it answers turns that are a choice among calls the robot already has, and it cannot write a joint angle. An LLM is still the pilot on every run, and --decision-mode shadow records what one would have picked without letting it change anything.

Question 3

What stops the model from picking something that breaks the robot?

Accepted Answer

A contract the model cannot talk its way out of: the goal and safety contract decides which skills a robot's model may use, how many steps it gets, when it must ask you first, and when to abort. A recorded 97-command SO-101 run fired no gate and kept every servo under 37 C, though the arm still fell at the end of every run because LeRobot drops torque on disconnect.

Question 4

Can I try it before buying any robot hardware?

Accepted Answer

Yes. Seven bodies are supported and six sit in a simulator or mock, including an Open Duck Mini you can print and build. The baseline is Python 3.11+ with uv, each robot is a separate install extra such as quackd[lerobot,openai], and quackd doctor checks the setup. Robot meshes are fetched under CC BY-NC-SA terms and never vendored.

Question 5

Which models can pilot a robot?

Accepted Answer

Claude, OpenAI, Gemini, Grok, Mistral, DeepSeek, Cohere, Qwen, Kimi, GLM and Meta work over their APIs, and OpenRouter-routed models behind one key. Open-source models run locally through Ollama, vLLM, llama.cpp or LM Studio with no key.

Question 6

How does a flock split work across robots?

Accepted Answer

Every robot in a flock gets its own model, and they divide the work by telling each other what they are going to do, so specialized robots can cooperate on one goal instead of one robot doing everything.

Question 7

What robots does it support?

Accepted Answer

Seven bodies: the Microduck biped, a printable Open Duck Mini v2, an SO-101 arm through LeRobot, a wheeled base over rosbridge, an XLeRobot dual-arm cart, an AlohaMini and a ToddlerBot humanoid, with one running on real hardware.

Quackd

About this project

Highlights

Quickstart

Questions about this entry

Reactions & coverage

From the Jev guides

From the Jev AI model community