←  Ian Zhang / Projects
Robot Learning / Post-training
Connecting —

PiPER
Pick‑and‑Place

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.

Robot
—
Policy
—
Data
—
Compute
—
Progress —
Estimated completion —
Step
—
 
Training loss
—
 
Best loss
—
 
Throughput
—
 
Epochs
—
 
Elapsed
—
 
01

Optimization

Training loss

L1 action error + KL regularizer, averaged per 200 steps

Vertical rules mark saved checkpoints.

Gradient norm

Global L2 norm before clipping

Learning rate

AdamW, transformer and ResNet-18 backbone

02

Throughput

Training speed

Steps per second — GPU compute vs. wall clock incl. checkpoint I/O

Time per step

Seconds, split into model update and data loading

Steps over time

Completed optimizer steps against wall-clock time, with projected finish

03

System

Free memory

System-wide, percent of 16 GB unified memory

Trainer memory

Resident set size of the training process, MB

Free disk

Gigabytes available for checkpoints

04

Checkpoints

StepSavedLossEpochPolicy sizeAvailability
The first checkpoint is written at step 5,000.
05

Configuration

Model
Training
Dataset