robo-jev
2B decision model for robots specialised in physical judgment, returning calibrated probabilities for questions like 'will the glass tip?'
Overview
robo-jev is a small decision model from Steerix Robotics that reads a short text of what a robot senses and answers in-the-moment physical questions, such as whether a load will tip if it brakes, whether a person reaches its path first, or whether a part seated, with a calibrated probability per answer in one forward pass. It is Qwen3.5-2B with a LoRA adapter trained from simulation in about an hour on one H100, and requests use the OpenRouter Decisions API shape (state plus 1-12 choice, noul or score questions).
On 1,100 held-out moments from its own simulator the README reports 0.80 physics skill versus 0.28 for TypeSafe Jev and near zero for other general decision models, ECE of 0.011 on yes/no, and much better option-order stability than the base. On an independent MuJoCo-labelled outside test the latest checkpoint gets risk direction right 83% of the time, close to Jev's 87%, but the README is explicit that it does not yet judge absolute risk levels beyond its simulator, and that general decision models remain ahead on general reading. It positions itself as System 1 above a System 0 control layer.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Yes / no | If I brake hard now, does the tall box tip? | — |
| Yes / no | Does the worker step into my lane before I can stop? | — |
| Choice | Which action is safer? | — |
How it works
- Describe the robot's sensed situation as short text
- Ask 1-12 typed physical-judgment questions
- Get calibrated probabilities in one forward pass
- Control layer (System 0) acts or calls a person
Strengths
- Large lead on its own physics benchmark per the README
- Well calibrated on yes/no (ECE 0.011 reported)
- Small and self-hostable
Caveats
- Absolute risk levels do not yet transfer outside its simulator
- Behind general models on general reading
- Main benchmark is its own
At a glance
- Type
- Project
- Status
- research
- Language
- Python
- Runtime
- PyTorch, self-hosted
- Has code
- yes
- License
- Apache-2.0
Use cases and integrations
Frequently asked questions
What does robo-jev decide?
Physical judgments for robots, such as whether something will tip, collide or seat, with calibrated probabilities.
How does it compare to Jev?
The README reports it far ahead on its own physics benchmark and close to Jev on an outside test of risk direction, but behind on general reading.
Does it control motors?
No; it is a judgment layer above the control layer.
Related projects
Open 4B System I driving model that scores behaviours like cruise, stop, yield and brake from three camera frames each decision.
ProjectresearchPythonMuJoCo PiPER arm controlled by a local decider-2b model choosing among ten skills via a Jev-style typed-choice interface while targets move.
Agent skillresearchPython- EmbodiedJev★ 302
Browser MuJoCo Franka Panda workbench comparing Jev, Claude, OpenAI-compatible and local models on pick, stack and obstacle tasks.
ServiceresearchPython Free hosted Jev-style multimodal decision model API (about 20B active) that picks actions from images and options, aimed at robot control.
ServicebetaRepository serving CLM-8B, a contrastive System One model behind a TypeSafe-compatible API, aimed at humanoid robot decisions.
ProjectresearchPythonIndependent open model on Molmo2-ER that answers spatial questions by choosing from fixed options with one shared classifier head.
ProjectresearchPython