Training loss
L1 action error + KL regularizer, averaged per 200 steps
An Action Chunking Transformer learning to pick up a block and place it on a target — trained from 50 demonstrations and live human corrections on a single Apple Silicon laptop.
L1 action error + KL regularizer, averaged per 200 steps
Global L2 norm before clipping
AdamW, transformer and ResNet-18 backbone
Steps per second — GPU compute vs. wall clock incl. checkpoint I/O
Seconds, split into model update and data loading
Completed optimizer steps against wall-clock time, with projected finish
System-wide, percent of 16 GB unified memory
Resident set size of the training process, MB
Gigabytes available for checkpoints
| Step | Saved | Loss | Epoch | Policy size | Availability |
|---|---|---|---|---|---|
| The first checkpoint is written at step 5,000. | |||||