Fine-tuning

  • finetuning
  • fine tuning
  • transfer learning
  • adapting a pretrained model
Definition

Fine-tuning starts from a checkpoint that already learned from other robots’ data and adapts it to your dataset, instead of starting from random weights. Four of the five models here are fine-tuned. ACT is the one trained from scratch, which is why it runs by far the longest schedule.

Last updated 2026-08-09

The two starting points, side by side

ModelStarts fromDefault stepsDefault learning rate
ACTRandom weights100,0001e-5
SmolVLAPretrained checkpoint20,0001e-4
GR00T N1.7Pretrained checkpoint20,0001e-4
GR00T N1.5Pretrained checkpoint2,0001e-5
Pi0.5Pretrained checkpoint30,0005e-5

The schedules carry the argument. Twenty thousand steps that adjust an existing network and one hundred thousand that build one are not the same kind of number, and the short schedule is not the cheap option: GR00T N1.7 finishes in 20,000 steps but needs an 80 GB card to do it, at 1.20 to 2.00 USD per hour, while ACT runs its 100,000 on a 24 GB card at 0.30 to 0.60 USD per hour.

What pretraining actually buys

  • Fewer episodes. SmolVLA sits at about 30 as a floor against about 50 for ACT, because it is not learning what a cube looks like from your data.
  • Tolerance for object positions you never demonstrated, which is the main argument for GR00T N1.7 once a dataset is already clean.
  • A language input that works on phrasings your episodes never contained, because the text side was trained on text rather than on your task string.
  • On GR00T N1.7 only a small part of the network moves during a fine-tune, roughly 40 million parameters out of about 3 billion.
Pretraining does not repair a dataset

A pretrained model handles unfamiliar scenes better than a small one trained from scratch. It does not handle contradictory demonstrations better, because that is a property of your data rather than of the network. If SmolVLA cannot learn the task at all on the defaults, moving to an 80 GB card usually buys a more expensive version of the same failure.

One thing stays yours either way: the normalization statistics computed from your dataset, and the calibration of the arm that recorded it. A fine-tuned checkpoint is a general model with your specific setup baked into the last mile, which is why it does not transfer to arbitrary hardware.

/policies compares the five recipes and says which one to reach for when, and /train has the arm-specific walkthroughs.