The day in conversation for 2026-10-23

The discussion: Rosa: That jump in the safe-success rate for SafeDojo is significant, 8.25 percentage points on Level I of SafeLIBERO—that’s not just incremental tuning; it feels like they’ve cracked the balancing act between ambitious policy learning and actual physical constraint adherence. Dev: It makes sense from a purely architectural standpoint, though I wonder if that world model imagination component is truly robust enough for deployment outside of highly controlled simulated environments. We’ve seen models get clever in simulation, but translating that internal imagination into reliable contact probability predictions on the real hardware remains the choke point. Taro: Exactly. And while we're admiring the safety gains, MiTaS achieving 80 percent success in contact-rich tasks is a different kind of breakthrough entirely—it’s about raw sensory integration and tactile feature fusion. That moves us away from just theoretical safety boundaries toward actual dexterity in messy, real-world manipulation scenarios. Rosa: But those two seem to be addressing fundamentally different problems. SafeDojo is optimizing the *risk* function during policy generation, whereas MiTaS is optimizing the *action* execution itself through superior perception of contact dynamics. I think the real headache lies in bridging that gap—how do we make a model that understands complex manipulation (MiTaS territory) inherently safe without needing those explicit, potentially brittle safety heads? Dev: That leads directly into why ReStruct is so intriguing, though it feels like a patch rather than a