Lesson 05 · 15 minutes · Preparation concept
Reasoning proposes.
Supervision decides.
Your next win: explain why a fluent visual-language response cannot directly move a robot.
Three jobs, three boundaries
| Component | Question it answers | May not do |
|---|---|---|
| Visual planner | What supported step fits the instruction and images? | Invent a motor command. |
| Action policy | What bounded joint targets fit this skill? | Skip limits or ownership. |
| Supervisor | Can this step run, recover, or stop? | Declare success without evidence. |
The supervisor owns prerequisites, timeouts, current arm ownership, cancellation, and recovery count. It validates a planner response against supported skills before a policy sees it. A clear stop with an explanation is a valid outcome when the scene or request cannot be handled safely.
Recovery needs a budget
fresh camera state→Try skill
bounded action→Check result
success or fault→Recover or stop
record outcome
The plan permits at most two recovery attempts for a failed step. Each attempt requires fresh observations; stale images can make a plausible retry dangerous or meaningless. Cancellation clears queued actions so an old plan cannot resume after a new instruction.
Practice: write a one-sentence operator explanation for “fork missing after two recovery attempts.” Include what was observed, what was tried, and why motion stopped. Ask the agent to critique it.
Primary reading: the project’s architecture and the Qwen3-VL model card for the proposed visual component.