NewJev as Robot Policy
The controller first asks Jev to select a manipulation intent from the task and observed geometry facts.
Added to Jevfast
NewHow it uses Jev
The controller first asks Jev to select a manipulation intent from the task and observed geometry facts. It then re-observes the arm, builds a target-relative state, and asks four Choice questions for x, y, z movement and finger position. Local code converts geometry into alignment/contact facts and turns the selected motor answers into controller actions.
What Jev decides
Illustrative source-derived request: choose manipulation intentExample answers · not a recorded Jev response · Source ↗
Question 1 · intent
YOUR APP
INSTRUCTION
Choose the next manipulation intent whose conditions match current facts.
STATE
Task name, geometry-derived observed facts, previous intent, and object height; the allowed intent criteria depend on task mode.
JEV · CHOICE
- approach
- grasp
- lift
- carry
- lower
- release
- withdraw
- finish
Illustrative source-derived request: choose axis directions and finger actionExample answers · not a recorded Jev response · Source ↗
Question 1 · x
YOUR APP
INSTRUCTION
Move along x toward the target or stay when aligned.
STATE
Selected intent, its definition, task name, current target-relative direction for each axis, geometry facts, and local target offsets.
JEV · CHOICE
- positive
- negative
- stay
Question 2 · y
YOUR APP
INSTRUCTION
Move along y toward the target or stay when aligned.
STATE
Selected intent, its definition, task name, current target-relative direction for each axis, geometry facts, and local target offsets.
JEV · CHOICE
- positive
- negative
- stay
Question 3 · z
YOUR APP
INSTRUCTION
Move along z toward the target or stay when aligned.
STATE
Selected intent, its definition, task name, current target-relative direction for each axis, geometry facts, and local target offsets.
JEV · CHOICE
- positive
- negative
- stay
Question 4 · fingers
YOUR APP
INSTRUCTION
Open, close, or keep the current finger opening based on intent and grasp-height facts.
STATE
Selected intent, its definition, task name, current target-relative direction for each axis, geometry facts, and local target offsets.
JEV · CHOICE
- open
- close
- stay
App workflow
- Derive observed facts
Local geometry code derives contact, alignment, height, and finger-state facts from simulator state.
- Select an intent
Jev selects among intent criteria that match current facts; lift-only task mode narrows and changes the criteria.
- Refresh observations
The arm may move during intent inference, so the code samples current geometry again before requesting motor choices.
- Choose motor controls
Jev returns axis and finger choices; local controller code performs the movement.
Why it is interesting
The design keeps geometry arithmetic and motion execution in the harness while Jev handles semantic intent and bounded motor choices. It refreshes observations between the dependent intent and motor requests.