LeRobot: Work Log
last updated
1. Introduction
The purpose of this article is to document my progression through different approaches to robot control and visuomotor policies using LeRobot and the SO-101.[1] The goal is to build an intuition for how robot learning has evolved: from classical model-based robotics, to reinforcement learning, to imitation learning, and finally to newer vision-language-action and world-model approaches. Along the way, I use small experiments in simulation and on the SO-101 to connect the theory to how these methods are actually used.
2. Classical Robotics
See Theory · Classical robotics enables precise, repeatable control for well-defined tasks by combining explicit models, motion planning, and feedback control. It works well when the robot and its environment are understood, but performance can degrade when reality differs from the model. Adapting to new tasks or operating conditions often requires redesigning or retuning the system.
3. Reinforcement-learning Policies
See Theory · Reinforcement learning lets robots learn control policies through interaction, including behaviours that are difficult to specify by hand. SAC reuses past experience to learn continuous control,[2] while HIL-SERL incorporates human demonstrations and corrections.[3] These methods can reduce manual controller design, but still depend on substantial experience and carefully designed rewards. Training in simulation helps with data collection, though transferring to a physical robot introduces a sim-to-real gap.
4. Imitation-learning Policies
Notes coming soon.
5. Vision-language-action Models (VLAs)
Notes coming soon.
6. World Models
Notes coming soon.