Online learning: sequential models and incremental architectures for real-time adaptation

A robot arm picks objects. While it works, four things need to keep learning, because the world will not stay the same:

• "How sure am I that this grasp will hold?" — the classifier that predicts "good grasp / bad grasp" was trained once, on a fixed set. If it says 0.8, that should mean 8 times out of 10. A model whose probabilities are not honest is dangerous on a robot: the control policy trusts them. → Calibration

• "Which past situations look like this one?" — the nearest-neighbour step underneath every retrieval, imitation-learning and anomaly-detection module needs a way to say "these two situations are similar". Euclidean distance is a default, not a decision: on real robot features it is often wrong. → Metric learning

• "Am I moving freely, touching something, or colliding?" — the robot's own state has to be recognised from its joint signals, sample by sample. If the contact signature changes (different payload, worn gripper), a classifier trained yesterday is stale today. → Online classifiers

• "How much torque does this move really need?" — the URDF model gives the rigid-body torque, but a real robot also has the payload in its gripper, friction in its joints, and wear. Those are not in the file; they have to be learned. → Online dynamics

The four share one property: they update one sample at a time, while the robot runs, without a separate training phase. That is what "online" means in this section.

Pages

• Calibration — online Platt scaling: corrects the probabilities of any classifier one sample at a time.
• Metric learning — OASIS, LEGO and POLA, and a k-NN classifier that learns its distance online.
• Online classifiers — robot-state classifiers that adapt when the data drifts.
• Online dynamics — learns, sample by sample, the torque the URDF model misses.