Polemic

The whole industry has a problem. There is a myopic obsession with reducing test loss. In the face of data walls and sim2real gaps, the field is grinding a rapid burn of capital with limited success in fundamentally unblocking new capabilities. You have to compress the dataset and task tune it; there is no other way.

The worst offender, robotics has learned to stage autonomy. Generalizable autonomy is nowhere close.

Of course, it doesn't have to be like this. Could you simply learn the policy during deployment time?

No no, you have to compress the dataset and task tune it. The herd warms itself by burning capital, best to not stray.

Catastrophic forgetting is seen as an immovable blocker to deployed online learning, so most seek incremental gains and complex scaffolds to collect scraps of progress.

However in the same industry, no one doubts deployed continual learning will be transformational in the limit. In our lifetime, continuous and lifelong learning will happen at deployment. This is a fact.

Infatuation with language models has put a chilling effect on the research that is required to support this, and the market needs a correction.

The path forward is clear, but it requires revisiting core primitives of neural computation. Brains orchestrate continuous learning through specific systems-level learning rules: fast episodic learning combined with slower consolidation.