Toward Agentic Robotics

Tracing the evolution from LLM-based planning through VLA to agentic robotics

What must change for the agentic code→tool→execution→error→LLM loop to work in the physical world? — 4 Parts, 10 Chapters

First published: 2026-04-08 | Last updated: 2026-04-12

Start Reading
🤖

Agentic Coding vs Agentic Robotics

Using digital agent success patterns as a mirror to analyze fundamental differences physical agentic systems must overcome.

📚

Research Evolution

Tracing paradigm shifts from LLM as Planner (2022) through VLA, Code as Policy to the latest Agentic Robotics.

🔬

50+ Core Papers Analyzed

Critical analysis of key papers including SayCan, RT-2, OpenVLA, pi0 with specific metrics and limitations.

Part I: Foundations — LLM Meets Robotics

01

Introduction — From Agentic Coding to Agentic Robotics

Core analogy, 7-dimension comparison, questions and structure

→ 02

LLM as Planner — Zero-Shot Planning and Grounding

LLM as Planners, SayCan, SayPlan, affordance grounding

→ 03

Code as Policies — Programming Robot Control

CaP, Code-as-Symbolic-Planner, CaP-X

→

Part II: The VLA Revolution

04

The Rise of Vision-Language-Action Models

PaLM-E, RT-2, OpenVLA, pi0

→ 05

Hierarchical Planning — From High-Level to Low-Level

AutoTAMP, Hi Robot, HAMSTER, RT-H

→ 06

Low-Level Control — Diffusion Policy and 3D Representations

Diffusion Policy, 3D Diffuser Actor, DROID

→

Part III: Toward Agentic Robotics

07

Memory and World Representation

KARMA, Embodied-RAG, Scene Graph

→ 08

Closed-Loop Agentic Systems

REFLECT, BUMBLE, AutoRT, PragmaBot

→ 09

Sim-to-Real Transfer and Evaluation

SIMPLER, Natural Language Sim2Real

→

Part IV: Fundamental Differences

10

Agentic Coding vs Agentic Robotics — The Gap and the Future

Comprehensive analysis, fundamental differences, future directions

→

Appendix

11

Appendix — Architecture of Agentic Coding Systems

Structural analysis of agentic coding systems

→

Appendices

A

Consolidated References

Full bibliography

→