← All projects

SO-101 VLA Imitation Learning — Finetuning SmolVLA

Abstract

This project walks through finetuning SmolVLA on the SO-101 arm: from hardware and camera setup, through demonstration recording and policy training, to a first deployment on the robot.

The goal is a repeatable VLA workflow on real hardware. I use a leader–follower SO-101 setup with wrist and top camera views, record imitation-learning demonstrations for a cup-stacking task, and train a SmolVLA policy in the LeRobot pipeline.

The video below shows that finetuned policy running on the physical arm. The robot was given the prompt “stack the cups” and executes the task from that instruction.