Evaluate end-to-end deployment latency and candidate batching

Evaluate the end-to-end deployment latency of ForkCast, including candidate generation and physical viewpoint acquisition, and determine the benefits of batching candidate evaluations for responsiveness in rapidly changing scenes.

Background

ForkCast’s reported runtime analysis measures computation from image loading through candidate scoring but excludes candidate generation, physical viewpoint acquisition or restoration, and execution. Consequently, the reported 1.439-second serial decision latency does not characterize the complete response time of a deployed robot system.

The authors identify two unresolved practical issues: the complete end-to-end latency remains unmeasured, and the potential advantages of evaluating candidates in batches rather than serially have not been established. These issues matter because additional sensing and computation may limit ForkCast’s responsiveness in rapidly changing scenes.