The day in conversation for 2026-11-09

discussion

Transcript

The discussion: Rosa: The takeaway from the conditioning signal work is that relying solely on history for robot policy uplink conditioning breaks down when motion becomes large; what does that imply about the robustness we saw in the trajectory planning module from last week's simulation runs? Dev: It suggests that for high-bandwidth scenarios, we need to prioritize global pose information over localized joint state sequences, but I’m interested in the implementation detail—how much computational overhead is this global pose tracking imposing on the actual embedded system compared to the gains we saw in policy success rates on OpenVLA? Taro: That computational load is a known trade-off, but I read about that INDI controller for powered descent; achieving 100% success across Monte Carlo campaigns means the control loop itself isn't introducing new failure modes related to phase transition instability, which is a big step from our previous iterations. Rosa: The 100% success in the INDI controller is impressive, but was that measured against scenarios involving external disturbances—for instance, unexpected wind gusts that might push the system outside the modeled uncertainty bounds? Because

More episodes

← Home