The robotics industry has found its comfort zone, and it's making me uncomfortable.
Watch the conversation around robot training right now. It orbits the same gravitational center: how do we get machines to learn faster, better, more reliably? Synthetic data, simulation scaling, reasoning-augmented training. These are real problems. The engineering is legitimate. But the consensus has calcified into something nearly religious: more learning capacity equals more robot capability equals progress.
The obvious takeaway, then, is that we need to invest harder in these methods. Build bigger datasets. Refine the simulations. Layer on reasoning systems. Rinse, repeat, celebrate.
The better question is what this obsession with learning breaks when it finally hits reality.
Consider the infrastructure assumption hiding inside this entire paradigm. We're training robots with the philosophy of abundance: abundant compute, abundant data, abundant correction cycles. This works in labs. It works in controlled environments where mistakes are expensive but not catastrophic. But what breaks when you deploy a robot into a space where none of those conditions hold?
What happens when the robot needs to forget?
I'm not talking about dramatic scenarios. I mean the mundane problem that nobody seems eager to address: a robot that learns from its environment learns everything from its environment. The noise, the anomalies, the one-off situations that should never have shaped its behavior in the first place.
A robot trained on thousands of simulated variations of a task might perform beautifully in controlled conditions. But the real world introduces variants the simulation never anticipated. The obvious move is to retrain, to feed that failure back into the learning pipeline. But what if the robot has already learned something from that environment that shouldn't have stuck around? What if the correction creates conflicts with earlier training?
The current consensus treats this as a data annotation problem. Label it correctly, weight it properly, and the system learns the right lesson. But this assumes learning is additive and always correctable, which is a comforting lie we tell ourselves in machine learning.
Biological systems have an answer: forgetting. Selective memory decay. Active pruning of irrelevant information. We don't fully understand these mechanisms, which is probably why robotics has sidestepped them entirely.
The industry has chosen the direction of maximum learning rather than optimal learning. That choice made sense when the bottleneck was data access and compute. But we're reaching a point where the constraint is shifting. It's becoming coherence.
A robot that learns everything becomes a robot that contains contradictions. A system trained on a thousand edge cases might perform well on average but fail unpredictably when multiple learned patterns intersect in unexpected ways.
This matters more as we scale deployment. Lab robots can be reset. Deployed robots can't. They inherit all their learning. If we've built systems that learn but not forget, we've built systems that accumulate inconsistency over time.
The consensus says this is solvable with better architectures, better training regimes, better reasoning systems layered on top. Maybe. But the industry's reluctance to even name forgetting as a design requirement suggests something else: we're still treating robots as pure learning machines instead of systems that need to actively manage their own knowledge.
The comfortable answer is that learning faster solves everything. The harder answer is that learning smarter might require unlearning. That's not where the funding flows. That's not where the papers get published.
But it's what breaks next.