Part III: Manufacturing Deployment

Chapter 7: Start at the Cell — Pick, Place, Inspect, Rework

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

The first question in Part III is not "which robot should we buy?" It is "which cell can we turn into a learning system?" NVIDIA's stack connects DGX, Omniverse, Isaac, Cosmos, GR00T, and Jetson into a fast experimentation loop, but the first durable win usually starts in a bounded pick/place/inspect/rework cell.

A good starting cell is one where human know-how, robot demonstrations, CAD or USD assets, quality images, and failure logs can be bound to the same task identifier. This chapter helps readers choose that first cell, decide what to log, and hand the result into the verification system of the next chapter.

Figure 7.1: Adoption depth by task primitive and operating maturity. illustration by author AI-assisted

Overview

Learning objectives - Break manual manufacturing work into pick, place, inspect, rework, and recover primitives. - Explain how contact, tolerance, lighting, and SKU variation change robot-policy difficulty. - Make a pilot produce a skill library, task schema, and failure dataset rather than only a demo video.

NVIDIA's physical-AI tools can accelerate iteration, but the manufacturer's advantage comes from preserving physical process experience as data. DeXtreme showed that policies trained in Isaac Gym can transfer to a real dexterous hand [1], while ForceVLA and Tactile-VLA style work shows why force and tactile signals matter for contact-rich work [2]; [5].

The first cell should therefore be a repeatable cell whose states and failures can be recorded, not a dramatic humanoid showcase. If that choice is right, sim-to-real validation, safety approval, and edge deployment can become one operating loop.

Cell type Good first candidate Candidate to avoid Data to preserve
Pick/place Repeated part, fixed fixture, clear grip pose Reflective material, thin film, flowing packaging Grasp pose, slip, drop, cycle time
Inspect Existing quality images and explicit pass/fail rules Human judgment that is implicit or frequently changing Pass/fail image, lighting, decision reason
Rework Three to five well-known failure modes Recovery behavior that varies widely by operator Failure class, recovery action, reinspection result
Insertion/assembly Measurable tolerance and force limits Deformable parts, cable routing, liquid, strict cleanroom rules Force trace, contact event, jam recovery

7.1 Choosing The First Cell

The first cell must satisfy both automation value and data collectability. A high-volume task is a weak candidate if its failures cannot be classified. A research-interesting task is also weak if cycle time, safety interlocks, and operator approval are unclear.

Use four criteria. The task primitive should be short and observable. The quality decision should be explicit. Failure should lead to a safe stop or human handoff. CAD/USD assets, camera logs, PLC events, teleoperation traces, and QA images should share the same task identifier.

The S6 manufacturing and logistics cases explain why many pilots stall after the first demo. Industry surveys show that firms can move beyond experimentation without seeing measurable value, while 2030 manufacturing forecasts emphasize staged adoption by use case and process rather than blanket automation [15]; [16]. The first cell should be not merely easy, but learnable.

7.2 Reading Work As Four Primitives

Pick looks like a grasping problem, but in production it combines lighting, pose estimation, gripper compliance, and slip detection. Place is stricter because the final pose affects fixtures, downstream work, and quality acceptance. Inspect is not just a vision model; it is a production decision. Rework is the circuit that turns failure into the next learning sample.

Figure 7.2: Manufacturing automation archetypes and cell-level priorities. source photo and illustration mix

Tactile work matters because all four primitives are connected by contact. Touch-only in-hand rotation shows that tactile and proprioceptive signals can serve as an independent perception channel [3]. PP-Tac shows why an apparently simple human action, such as picking one sheet of paper, can require tactile policy learning for robust robotic execution [4].

Manufacturers should not read this as "hands will eventually become good enough." The immediate task is to decide which signals each primitive requires. Pick needs slip and grip force; place needs pose error and fixture contact; inspect needs image lineage and decision reasons; rework needs a failure taxonomy and records of human intervention.

7.3 Build The Skill Library First

A skill library is not just a code repository. It is a process-knowledge repository. A pick skill for a bottle, cap, tray, cable, or thin film can require different limits and recovery paths. Each skill should include input state, allowed action, force limit, forbidden state, quality rule, and recovery behavior.

Demonstration data needs the same discipline. ALOHA and Mobile ALOHA show how low-cost teleoperation and imitation learning can expand bimanual manipulation data collection [12]; [11]. In a factory cell, however, the key fact is not only how many demonstrations exist. The key fact is which SKU, fixture, operator, and process condition produced each demonstration.

Start with a task schema. At minimum, track task_id, sku, fixture_version, robot_config, sensor_config, operator_mode, quality_rule, failure_class, and policy_version. This schema lets Isaac simulation and real-cell logs meet under the same evaluation truth.

7.4 Manufacturing Cell Checkpoint

Checkpoint Question Passing condition
Task boundary Are start and end states clear? The cell can be labeled pass/fail without ad hoc discussion
Safety stop Where does the system stop after failure? Robot, PLC, and operator override paths are documented
Data lineage Under what condition was the data captured? Demonstrations, simulation, and QA images share a task id
Retry policy What is retried after failure? Retry count, regrasp, and human handoff rules exist
Learning handoff Does failure enter the next experiment? Failure class flows into synthetic data or evaluation cases

If a cell does not pass this checkpoint, even robot success leaves behind a weak asset. If it passes, a bounded cell becomes the manufacturer's first real skill library.

7.5 What To Learn Next

The cell selected here moves into the verification and safety system of Chapter 8. The next learning goal is not another model architecture. It is a release ladder: shadow mode, supervised mode, restricted production, and fleet rollout, each with a clear evidence package.

Concretely, the team should reproduce failure cases in Isaac or another simulator, measure how often the same failures appear in the physical cell, and connect operator intervention plus quality decisions back into model improvement. That is what turns Part III from an adoption guide into a production-responsibility framework.

References

  1. Ankur Handa et al. (2023). DeXtreme: Transfer of Agile In-hand Manipulation from Simulation to Reality. ICRA 2023. https://arxiv.org/abs/2210.13702
  2. Jiawen Yu et al. (2025). ForceVLA: Enhancing VLA Models with a Force-aware MoE for Contact-rich Manipulation. NeurIPS 2025. https://arxiv.org/abs/2505.22159
  3. Z.-H. Yin et al. (2023). Rotating without Seeing: Towards In-hand Dexterity through Touch. RSS 2023. https://arxiv.org/abs/2303.10880
  4. Pei Lin et al. (2025). PP-Tac: Paper Picking Using Omnidirectional Tactile Feedback in Dexterous Robotic Hands. RSS 2025. https://arxiv.org/abs/2504.16649
  5. Jialei Huang et al. (2025). Tactile-VLA: Unlocking Vision-Language-Action Model's Physical Knowledge for Tactile Generalization. arXiv preprint arXiv:2507.09160. https://arxiv.org/abs/2507.09160
  6. Jessica Yin et al. (2024). Learning In-Hand Translation Using Tactile Skin With Shear and Normal Force Sensing. arXiv preprint arXiv:2407.07885. https://arxiv.org/abs/2407.07885
  7. Erik Helmut et al. (2025). Tactile-Conditioned Diffusion Policy for Force-Aware Robotic Manipulation (FARM). ICRA 2026. https://arxiv.org/abs/2510.13324
  8. Peng Hao et al. (2025). Tactile-Language-Action Model for Contact-Rich Manipulation (TLA). arXiv preprint. https://arxiv.org/abs/2503.08548
  9. Cheng Chi et al. (2023). Diffusion Policy: Visuomotor Policy Learning via Action Diffusion. Robotics: Science and Systems (RSS) 2023 / IJRR 2024. https://arxiv.org/abs/2303.04137
  10. OpenAI et al. (2020). OpenAI Dactyl: Solving Rubik's Cube with a robot hand. arXiv 1910.07113 / OpenAI tech report. https://arxiv.org/abs/1910.07113
  11. Zipeng Fu et al. (2024). Mobile ALOHA: Learning Bimanual Mobile Manipulation with Low-Cost Whole-Body Teleoperation. arXiv preprint. https://arxiv.org/abs/2401.02117
  12. Tony Z. Zhao et al. (2023). Learning Fine-Grained Bimanual Manipulation with Low-Cost Hardware (ACT/ALOHA). Robotics: Science and Systems (RSS) 2023. https://arxiv.org/abs/2304.13705
  13. Claire Chen et al. (2025). DexForce: Extracting Force-informed Actions from Kinesthetic Demonstrations for Dexterous Manipulation. IEEE Robotics and Automation Letters. https://arxiv.org/abs/2501.10356
  14. Miquel Oller et al. (2024). Tactile-Driven Non-Prehensile Object Manipulation via Extrinsic Contact Mode Control. Robotics: Science and Systems (RSS) 2024. https://www.roboticsproceedings.org/rss20/p135.pdf
  15. BCG and Alpega (2026). AI Is Already Moving the Logistics Industry Forward. Boston Consulting Group.
  16. PwC (2026). Industrial Manufacturing's Race to 2030. PwC Global Industrial Manufacturing Sector Outlook.