Design note: the autonomous-intelligence architecture and Shappy

The goal is a robot with a persistent self that walks, talks and carries out varied tasks. LeCun's 2022 architecture splits that into a configurator, perception, world model, cost, actor and memory; the walking policy is the reactive actor and the reward is part of the cost, while the JEPA world model, its encoder, the memory and the dialogue link are not built.

• Shappy's walking policy (50 Hz) corresponds to the reactive actor; planning speed commands through a world model would be the second actor mode and does not exist yet
• The detection recordings made so far cannot train a world model: boxes and a timer are burned into the frames, there are no action, joint or IMU records beside them, and most runs follow one person
• Recording to build: plain frames at about 10 per second, joint targets and speed commands at the control rate, joint and IMU states, policy file and run notes, on one clock; driven sessions and slippery or obstructed ground included. Estimated 1–2 GB per hour
• Order: check frame–action timing on a short session, test a pretrained JEPA encoder on the fisheye footage before collecting more, train the predictor on skim and use it first as a surprise detector
• Short-term memory in the paper is a working memory of current and predicted states and their costs, not a dataset: the recordings on skim train the world model, and the memory itself would be a runtime buffer on the robot. The paper's Fig. 2 is reproduced in the report for reference