Latency-Aware Robot Policy Runtime
Freshness-aware scheduling for asynchronous learned control
I built an asynchronous learned-policy runtime that tracks every action back to its source observation and compares synchronous, FIFO, and latest-only scheduling semantics.
Impact
On MuJoCo Reacher with 100 ms policy latency, FIFO and LATEST both maintained about 20.4 Hz control. FIFO nevertheless accumulated 1.24 s p95 action age and reached only 20% success. LATEST reduced p95 action age to 150 ms and restored 100% success while keeping the same control rate.
The project demonstrates why throughput alone is not a sufficient deployment metric for closed-loop robot policies.