Sign InOpen Brain
arXivPaperNeeds Review

RAPID: Robot Agentic Programming from Demonstrations

RAPID turns one visual demonstration into an executable, testable robot program, showing how agent loops can derive their own specifications, tools, and verification environment.

arXiv · Sep 24, 2026
Open Source Open MarkdownOpen JSON
Source Summary

RAPID infers a task specification, robot action primitives, and an execution environment from **one visual demonstration**, then generates, verifies, and refines a program.

Practical Implication

The reusable pattern for coding-agent builders is the closed loop: derive executable tests and tools from an example, run the program, and refine it against observed behavior. Its object-centric representation also separates strategy from scene-specific geometry.

Agent-Ready Context
RAPID infers a task specification, robot action primitives, and an execution environment from **one visual demonstration**, then generates, verifies, and refines a program.

The reusable pattern for coding-agent builders is the closed loop: derive executable tests and tools from an example, run the program, and refine it against observed behavior. Its object-centric representation also separates strategy from scene-specific geometry.

The evidence covers simulation and a Franka arm across **eight contact-rich tasks**, plus LIBERO-Pro evaluations. The abstract reports generalization across several variations but provides no comparative rates.
Connected Context · Feed7 Judgment

RAPID extends example-driven agent programming into physical control: one demonstration can seed not only a program but its task specification, action vocabulary, execution environment, and verification loop. The object-centric split between strategy and geometry provides a concrete route to reuse across scene variations, though eight tasks and unquantified comparative results leave the breadth of that reuse unsettled.

Context Map
agentcodingimage#coding-agents#harness-engineering#tool-use
Uncertainty
The evidence covers simulation and a Franka arm across **eight contact-rich tasks**, plus LIBERO-Pro evaluations. The abstract reports generalization across several variations but provides no comparative rates.