Robot Motion 101 (3/3)
From VLA and World Models to Agentic Robotics
A research-grade practical guide from robot foundation data through VLA, world models, agents, and verified deployment — 3 Parts, 9 Chapters
First published: 2026-07-16 | Last updated: 2026-07-16
One Execution Contract Across Three Volumes
Connect natural-language intent to skill proposals, feasibility projection, real-time control, and independent safety authority.
From VLA to World Models and Agents
Compare data, action representations, adaptation, grounding, prediction, long-horizon planning, and skill APIs on common axes.
Verifiability Before Generality
Separate contamination, failure, latency, recovery, and operating cost to gate bounded deployment and exact rollback.
Part I: Understand the Materials and Structure of Robot Foundation Models
Where Does Higher-Level Intelligence Connect? — The Task, Skill, and Control Contract Across Three Volumes
Fix the proposal boundary and independent safety authority for VLA, world models, and agents atop the S11 cell and S12 execution spine.
→ 02What Do We Pretrain On? — Robot Data, Cross-Embodiment Learning, and Action Representations
Compare robot, video, language, and synthetic data plus continuous, chunked, tokenized, latent, and skill-call actions through provenance and embodiment mappings.
→ 03How Does a VLA Produce Actions? — Model Families, Training Objectives, and Policy Heads
Synthesize multimodal fusion and action-head designs on common axes of I/O, horizon, closed-loop execution, and real-robot evidence.
→Part II: Adapt to New Robots, New Tasks, and Long Horizons
How Do We Fit a New Robot and Task? — Fine-Tuning, Action Heads, and Embodiment Adaptation
Separate full and partial tuning, adapters, new action heads, distillation, and retrieval, then produce a reversible adaptation receipt.
→ 05Ground Language and Scenes into Action — Open Vocabulary, Spatial Understanding, and Skill Selection
Translate instructions into task constraints, goal states, and skill preconditions while testing rejection of ambiguity, conflict, and hidden state.
→ 06Can Prediction Support Planning? — World Models, Video/Action Prediction, and Policy Coupling
Separate latent, video, action, reward, and uncertainty prediction and gate model outputs as candidates requiring independent verification.
→Part III: Extend to Agents Without Losing Verifiability
What Should a Robot Agent Own? — Planning, Memory, Tools, and Skill APIs
Split long-horizon decomposition, skill calls, memory, tools, code generation, and human approval into separate authorities and recovery states.
→ 08How Can We Trust Generality? — Evaluation, Contamination, Failure, Safety, and Operations
Separate generalization regimes and record interventions, unsafe proposals, rejection, latency, recovery, shift, and remote operations in an independent evaluation card.
→ 09Complete the Three Volumes as One Robot System — From Instruction to Verified Execution
Compare the classical baseline, S12 policy, VLA, optional world model, and agent on tabletop assembly under fixed execution, evaluation, and safety contracts.
→