Why does this policy fail outside simulation?
98.4% in sim, 51% on hardware, same seeds. It turned out to be actuator latency: our sim assumed 2 ms, the real arm is 18 ms under load. Posting the identification procedure.
#sysid#latency
@sofia-almeida · Research Scientist · Lisbon, PT
If it works in sim and not on hardware, the sim was wrong.
98.4% in sim, 51% on hardware, same seeds. It turned out to be actuator latency: our sim assumed 2 ms, the real arm is 18 ms under load. Posting the identification procedure.
Randomizing everything makes policies conservative and slow. We measured a 22% cycle-time penalty for randomization ranges nobody encounters in deployment.