Full system prompt revealed: how LLMs control a real robot arm via tool calls

YuXiang_IRVL · x · 2026-09-06

YuXiangIRVL dug out the full system prompt used to control a real robot embodiment (yamarms) via tool calls. Key design: each observation gives proprioceptive state and camera images; the model moves in small deliberate steps and re-checks observation after each motion; every move must include a human-readable note explaining what it sees and why it chose that motion; a safety approver layer clamps out-of-bounds and too-fast actions; mid-run operator feedback is treated as trusted human guidance; exactly one tool call per turn. A practical look at LLM teleoperation engineering: observation-decision loops, transparent intent notes, safety clamping, and human supervision.

Original post →

More from coding & agent

coding & agent channel →