Chapter 7: Start at the Cell — Pick, Place, Inspect, Rework
The right starting point is not a humanoid purchase; it is a cell-level task decomposition. Manual work should be decomposed into pick, place, inspect, rework, load, unload, wipe, insert, fasten, and align primitives. Each primitive should then be classified by visual observability, force/tactile dependence, hygiene risk, and line-stop cost.
7.1 Close the Loop on Easier Tasks First
Initial targets should use standard containers and clear failure criteria: tray loading, bottle alignment, label inspection, reject-bin sorting, box loading, and palletizing. These can be simulated in Isaac, augmented with synthetic trajectories, deployed on edge compute, and evaluated through existing inspection outputs.
7.2 Build a Data-Collection Cell
A practical first step is a data-collection cell. Workers perform the task normally while wrist cameras, external cameras, force-torque sensors, and inspection results are recorded. UMI and DexUMI-style approaches turn this human work into robot learning data [1] [2].
References
- Cheng Chi et al. (2024). Universal Manipulation Interface. arXiv.
- DexUMI Team (2025). DexUMI. arXiv.
- Amazon Robotics (2025). Amazon Robotics and AI in Fulfillment. Amazon News.