The day in conversation for 2026-10-13

The discussion: Dev: I’m really stuck on this idea from w squared VLA about decoupling conceptual understanding from physical action. If you can truly separate those two layers, zero-shot transfer shouldn't be such a massive hurdle anymore. It implies we don't need massive datasets of specific demonstrations for every new task; we just need the foundational concept mapping to hold up. Rosa: I see the potential there, Dev, but I worry that separating them perfectly is an idealized goal. In practice, those layers are inherently entangled in current state-of-the-art models because they are all optimized end-to-end for reward maximization on specific tasks. We might get better *proxies* for decoupling rather than a clean break, but the pursuit of that separation is clearly where the future of general robotics lies. Taro: If we look at it through the lens of representation, DeFM seems to offer a more grounded approach than just trying to force VLA models to learn this structure. Learning robust geometric and semantic representations directly from depth images bypasses some of the ambiguity in how vision models interpret abstract language instructions for physical manipulation. That distillation aspect is what makes it compelling—taking a huge pre-training model and shrinking its complexity while retaining that geometric fidelity. Dev: I agree DeFM is a powerful foundation, especially since we spent last week talking about how brittle those purely visual policies are when the environment shifts slightly. But then you have FIDE L stepping in with the failure detection module. That’s where I think the real engineering challenge lies: not just making it *work* well on one task, but making it *detect* when it breaks in the messy real world. Rosa: FIDE L feels like a necessary sanity check, though. It moves us from purely aspirational learning—"can the model do this?"—to practical deployment—"is this specific instance of doing this actually viable given the sensor noise or hardware wear?" If we can reliably flag failure modes, then perhaps we don't need perfect zero-shot transfer anymore; we just need robust systems that