Large-Data Manipulation
Data engines, control, and learning for manufacturing automation
How data flywheels are reshaping manufacturing manipulation. — 3 Parts, 12 Chapters
First published: 2026-06-18 | Last updated: 2026-06-18
Manufacturing First
Frames manipulation through factory cells, quality gates, failure logs, and repeatable work rather than demos alone.
Data Flywheels
Explains how human demonstrations, robot fleets, simulation, force/touch, and QA traces become one learning loop.
Research and Startups
Maps the Tedrake, Finn, Abbeel, and Levine lineages alongside US startup strategies.
Part I: Manufacturing Manipulation and the Data Bottleneck
Why Manufacturing Manipulation Is a Large-Data Problem
Defines high-mix, high-variance, contact-rich work as a data problem.
→ 02Choosing the Hand: Two Fingers, Suction, Custom Hands, and Five Fingers
Compares how end-effectors reshape data collection and learning difficulty.
→ 03The Data Flywheel: Humans, Robots, Simulation, and Failure Logs
Defines the task data and evaluation harness manufacturers must own.
→Part II: Research Lineages in Control and Learning
The Tedrake Lineage: Contact, Drake, and Sim-to-Real
Explains how model-based control and contact modeling support large-data strategies.
→ 05The Abbeel, Levine, and Finn Lineage: Imitation and RL
Connects robot data collection, imitation learning, reinforcement learning, and offline learning.
→ 06Policy Architectures: ACT, Diffusion Policy, RT-X, OpenVLA, and pi0
Maps policy architectures that include VLAs without reducing the field to VLAs.
→ 07From Humans to Robots: UMI, DexUMI, DEXOP, and EgoScale
Compares strategies for turning human hand and work data into robot-executable data.
→ 08Contact-Rich Work: Touch, Force, and Online Improvement
Explains why vision-only scaling needs tactile and force-rich data.
→Part III: Data Strategies Through Companies
US Foundation-Model Labs: PI, Generalist, Skild, and Figure
Compares the data strategies of US startups building general physical AI models.
→ 10Production Flywheels: Covariant, Dexterity, and Chef Robotics
Analyzes how deployed production data becomes model improvement and deployment moat.
→ 11Hand and Hardware Co-Design: Sunday, Eka, Genesis, and Sanctuary
Compares data gloves, custom hands, five-finger hands, and touch/force integration.
→ 12What Manufacturers Must Own
Uses Config, CarbonSix, and AgiBot as supporting cases for buy/build decisions.
→