Josef Liyanjun Chen
This paper proposes a new economic framework that models a robot's limited flash memory wear as depreciating capital to determine optimal memory placement across different storage tiers.
Robots use multiple memory hierarchies (RAM, on-board NVM, cloud), but flash memory has a finite and irreversible number of program/erase cycles. There is currently no systematic way to decide which memories are worth consuming this scarce endurance, hindering cost-efficient memory management.
The authors treat flash endurance as depreciating capital and introduce a single endurance shadow price, η. This price becomes the threshold in a wear-augmented per-byte index that dictates cost-minimizing placement across RAM, NVM, and cloud. They also measure the value-write association (χ) on real robot logs to analyze how it varies with deployment regimes (long-horizon manipulation, short-horizon tasks, teleoperation).
The measured χ values vary by deployment regime: positive for long-horizon manipulation, null for short-horizon tasks, and negative for non-recurrent teleoperation. This shows that optimal memory placement can be non-monotonic. The study also finds that while the endurance budget is dormant on premium TLC flash, it becomes binding on the commodity QLC/eMMC used in cheaper edge robots. Furthermore, a learned wear-aware controller only ties price-based routing on device lifetime and cost, not on task performance.