This is a starter post — replace it with your own writing. It exists so the homepage Writing section and the
/posts/archive render with real content.
Every robotics demo looks like perception. A camera, a neural net, a crisp bounding box. But the demo that matters is the one that closes the loop: perception that commits to a decision, and a decision that survives contact with the real world.
The loop, not the layer
It’s tempting to treat perception, planning, and control as a pipeline — clean handoffs between specialists. In contact-rich manipulation that abstraction leaks immediately. The moment the gripper touches the part, perception and action stop being separate stages and become a single feedback loop running at the speed of physics.
Where the bottleneck actually is
A few things I keep coming back to:
- Sample efficiency. World models earn their keep here — imagining outcomes is cheaper than collecting them on real hardware.
- Grounding. Vision-language-action models can map instructions to behaviour, but “grounding” is doing a lot of work in that sentence.
- Memory. Long-horizon tasks need spatial and episodic memory, not just a bigger context window.
More to come.