Part IV: Manufacturing Strategy

Chapter 10: What to Buy and What to Build — Platform Dependency and Data Assets

Written: 2026-06-08 Last updated: 2026-06-11

The core question in Part IV is not technology choice. It is ownership choice. NVIDIA and its partner ecosystem can help manufacturers start training, simulation, modeling, and edge deployment quickly. The long-term assets a manufacturer must own are process data, failure logs, evaluation criteria, and production-approval authority.

A good strategy does not avoid vendor dependency entirely. It separates the layers where dependency is acceptable from the layers that must remain internal. This chapter provides a practical frame for what to buy, what to build, and what to co-develop.

Figure 10.1: ROI and automation difficulty matrix for consumer-goods manufacturing. illustration by author AI-assisted

Overview

Learning objectives - Separate purchasable general infrastructure from process knowledge the manufacturer must own. - Design ownership for robot policies, teleoperation data, digital twins, edge runtimes, and quality gates. - Define internal capabilities that survive vendor roadmap changes.

Foundation models such as GR00T, OpenVLA, and pi0.5 are expanding the common base for general robot intelligence [1]; [2]; [11]. In manufacturing, however, the durable moat is less the model itself than process-specific data exhaust: which part slips when, which fixture jams on which SKU, and which rework action restores quality.

The strategy question is therefore closer to "what data and evaluation system will we accumulate?" than to "will we build the foundation model ourselves?"

Layer Reasonable to buy Must own internally Good co-development target
Foundation model VLA backbone, humanoid policy base Process-specific fine-tuning and evaluation set Contact-rich skill adaptation
Simulation Omniverse/Isaac toolchain, partner plugins USD assets, fixtures, tolerance model Physics calibration, synthetic scene generator
Data capture Teleop hardware, mocap, sensors Demonstration lineage, operator consent, QA labels Human-hand interface, exoskeleton capture
Edge runtime Jetson/IGX stack, safety-certified components Deployment policy, rollback rule, local logs Model compression, latency tuning
Governance Vendor model card, reference architecture Release gate, audit trail, quality approval External audit, safety-case template

10.1 Buy The General Layers That Create Speed

Manufacturers do not need to build everything themselves. GPUs, simulation engines, robot middleware, reference models, and edge inference stacks are often faster and safer to buy. Industrial partnerships such as ABB RobotStudio HyperReality and Rockwell's edge AI work with NVIDIA point in this direction [13]; [14].

The criterion for buying is generality. Functions reused across many processes and plants, functions with heavy security or maintenance requirements, and functions where ecosystem compatibility matters are usually better as platform purchases.

Procurement should still specify data and update rights. The manufacturer should prevent raw data, derived features, failure taxonomies, and quality labels from becoming trapped inside vendor lock-in.

10.2 Build The Process Experience And Evaluation Truth

The manufacturer should build the physical experience of its own process. DexUMI, AnyTeleop, DexCap, and ExoStart show several ways to capture human demonstrations and dexterous manipulation data [5]; [3]; [8]; [9]. Manufacturers should absorb these methods into their own process data structures.

Data assets are not merely video and sensor logs. They include quality-decision reasons, operator overrides, fixture changes, cleaning cycles, shift conditions, material lots, and rework success. Without this context, a demonstration can be a learning sample but not production knowledge.

Figure 10.2: Strategic balance between proprietary process data and platform dependency. illustration by author AI-assisted

Evaluation truth must also be owned internally. If the model changes, the same test suite should compare old and new behavior. If the robot hardware changes, task success, defect rate, safety stop, cycle time, and operator burden should still be measured with a common standard.

10.3 Co-Develop Contact-Rich Skills And Field Integration

Co-development is most useful at boundary layers. Insertion, wiping, screwing, flexible packaging, and cable handling require both vendor model/tooling expertise and the manufacturer's process knowledge.

ManipTrans, HAMSTER, and Hi Robot point toward hierarchical and retargeting approaches for broader embodiment and task generalization [6]; [7]; [10]. Manufacturers should read this not as a promise that one general robot will do everything, but as an opportunity to turn process-specific skills into reusable interfaces.

The core issue in co-development is ownership of outputs. The parties should specify who may reuse fine-tuned weights, synthetic datasets, calibrated USD assets, failure suites, and edge deployment profiles.

10.4 Manufacturing Cell Checkpoint

Checkpoint Question Passing condition
Data ownership Are raw and derived data rights clear? Contracts and internal policy use the same definitions
Evaluation portability Does evaluation survive vendor replacement? Test suites are versioned separately from models
Skill interface Is the skill tied to one robot? Inputs, outputs, force limits, and failure classes are abstracted
Update control Who approves vendor model updates? Production deployment must pass the internal release gate
Exit path What if platform dependency becomes excessive? Data export, simulator assets, and edge fallback plans exist

This checkpoint is not meant to slow purchasing. It lets the company buy quickly while preserving the structure needed to learn.

10.5 What To Learn Next

The next chapter turns ownership choices into a 2026-2030 roadmap. The sequence is to build the first cell and data structure in 2026, harden simulation and validation in 2027, and expand into multi-cell operation and a production operating system from 2028 to 2030.

Before reading it, write three sentences for your organization: "We buy..." "We own..." "We do not deploy without this evidence..." Those sentences are the starting point for manufacturing physical-AI strategy.

References

  1. Johan Bjorck et al. (2025). GR00T N1: An Open Foundation Model for Generalist Humanoid Robots. arXiv preprint. https://arxiv.org/abs/2503.14734
  2. Moo Jin Kim et al. (2024). OpenVLA: An Open-Source Vision-Language-Action Model. arXiv preprint. https://arxiv.org/abs/2406.09246
  3. Yuzhe Qin et al. (2023). AnyTeleop: A General Vision-Based Dexterous Robot Arm-Hand Teleoperation System. Robotics: Science and Systems (RSS) 2023. https://www.roboticsproceedings.org/rss19/p015.pdf
  4. Prithwish Dan et al. (2025). X-Sim: Cross-Embodiment Learning via Real-to-Sim-to-Real. arXiv preprint. https://arxiv.org/abs/2505.07096
  5. Mengda Xu et al. (2025). DexUMI: Using Human Hand as the Universal Manipulation Interface for Dexterous Manipulation. CoRL 2025 (Best Paper Finalist). https://arxiv.org/abs/2505.21864
  6. Kailin Lv et al. (2025). ManipTrans: Efficient bimanual dexterous manipulation retargeting. CVPR 2025. https://arxiv.org/abs/2503.21860
  7. Yi Li et al. (2025). HAMSTER: Hierarchical Action Models For Open-World Robot Manipulation. arXiv preprint. https://arxiv.org/abs/2502.05485
  8. Chen Wang et al. (2025). DexCap: Scalable and Portable Mocap Data Collection for Dexterous Manipulation. arXiv / RSS. https://dex-cap.github.io/
  9. Zilin Si et al. (2025). ExoStart: From 10 Exoskeleton Demos to Dexterous Robot Manipulation. arXiv preprint. https://arxiv.org/abs/2506.11775
  10. Lucy Xiaoyang Shi et al. (2025). Hi Robot: Open-Ended Instruction Following with Hierarchical Vision-Language-Action Models. arXiv preprint. https://arxiv.org/abs/2502.19417
  11. Physical Intelligence (2025). pi0.5: A Vision-Language-Action Model with Open-World Generalization. arXiv preprint. https://arxiv.org/abs/2504.16054
  12. Hao-Shu Fang et al. (2025). DEXOP: Passive Exoskeleton for Direct-contact Dexterous Demonstration. arXiv preprint. https://arxiv.org/abs/2509.04441
  13. ABB Robotics and NVIDIA (2026). Closing the Sim-to-Real Gap: ABB RobotStudio HyperReality Enables Industrial-Scale Physical AI. ABB Press Release.
  14. Rockwell Automation and NVIDIA (2025). Rockwell Automation to Advance Industrial Intelligence Through Edge-Based Generative AI with NVIDIA Nemotron. Rockwell Automation Press Release.