People ask me what to read to get up to speed on Physical AI. I never send them the latest paper. I send them the old ones… the questions were asked decades ago, mostly by people looking at brains and at stubborn robots, and the new work makes a lot more sense once you know that.
So here are twelve papers, in the order they were written. Only one is mine. For each, a line on why I keep going back to it.
Last updated: 12 November 2026.
The twelve papers
1987: A massively parallel architecture for a self-organizing neural pattern recognition machine
Carpenter & Grossberg, Computer Vision, Graphics, and Image Processing. The first full Adaptive Resonance Theory network. The question it asks is still the one I care about most: how does a system keep learning new things without erasing the old ones? The answer here is to learn only when the input matches what the network expects. Read the paper.
1989: Catastrophic Interference in Connectionist Networks: The Sequential Learning Problem
McCloskey & Cohen, Psychology of Learning and Motivation. Two psychologists teach a network its additions one set at a time, and watch the earlier ones fall apart. The paper gave the problem its name. Thirty-seven years later it is still on the to-do list. Read the paper.
1990: Neuromorphic electronic systems
Mead, Proceedings of the IEEE. Carver Mead makes the case for building electronics the way nervous systems are built. Read it for the argument about energy… it has aged very well. Read the paper.
1991: Intelligence without representation
Brooks, Artificial Intelligence. Rodney Brooks argues that you get to intelligence by building complete creatures that act in the real world, one simple layer of behavior at a time. Half the field was annoyed. The robots worked. Read the paper.
1995: Why there are complementary learning systems in the hippocampus and neocortex
McClelland, McNaughton & O’Reilly, Psychological Review. Why does the brain have a fast learner and a slow one? Because one system cannot do both jobs without tripping over itself. This is the biology behind what engineers now call replay. Read the paper.
1999: Predictive coding in the visual cortex
Rao & Ballard, Nature Neuroscience. Higher visual areas predict what lower areas should be seeing, and only the errors travel upward. If you want one paper on the brain as a prediction machine, this is the one I hand out. Read the paper.
2008: Spikes, synchrony, and attentive learning by laminar thalamocortical circuits
Grossberg & Versace, Brain Research. The one I co-wrote, so discount accordingly. It is a model of how the layers of the cortex and the thalamus work together to decide when to learn and when to hold still. Long, and not a beach read. Read the paper.
2017: Overcoming catastrophic forgetting in neural networks
Kirkpatrick et al., PNAS. The paper that put forgetting back on the agenda of deep learning. The idea: find the connections that mattered for the old task and make them harder to change. Read the paper.
2017: Domain Randomization for Transferring Deep Neural Networks from Simulation to the Real World
Tobin et al., IROS 2017. A lovely trick. Make the simulator look wrong in thousands of different ways (colors, lights, textures) and the real world becomes just one more variation. Read the paper.
2018: World Models
Ha & Schmidhuber, arXiv. An agent learns a small model of its world and then practices inside it. Brains rehearse too. Short, clear, and with pictures. Read the paper.
2023: RT-2: Vision-Language-Action Models Transfer Web Knowledge to Robotic Control
Brohan et al., arXiv. Take a model trained on images and text from the web and have it output robot actions. This is where “vision-language-action” became the recipe everybody talks about. Read the paper.
2023: Open X-Embodiment: Robotic Learning Datasets and RT-X Models
Open X-Embodiment Collaboration, arXiv. Many labs pool the data from their very different robots and train one model on all of it. The interesting result is that robots can learn from other robots’ experience. Read the paper.
In what order?
If you have one evening: Brooks (1991), then McCloskey and Cohen (1989), then RT-2 (2023). You get the body, the memory problem and where the field is today.
If you have a weekend, read them all in order. You will notice that the list goes from neuroscience and psychology journals to robotics conferences. That is the history of the field in one glance.
What is missing
A lot. Twelve is a small number, and this is a personal list. If you think a paper belongs here, tell me and tell me which one it should replace.