Skill Segmentation: the right control method for every subtask

At Servo7, we build robots in the field every day. Along the way, we build tools to solve the problems we keep running into. Skill segmentation is one of them. We are sharing it because we are tired of seeing other robotics teams solve the same problems from scratch. Roboticists should spend their time on hard problems, not reinventing tooling.

The problem

Take a single task: picking an onion from a bin and placing it in a mealbox. Simple on paper. In practice it consists of four distinct subtasks: grasping the onion, moving with it from A to B, placing it at B, and moving back to the start position.

Each of those subtasks has different properties. Grasping requires precision and has to adapt to variation in how the onion sits. The transit from A to B is fast and relatively unconstrained. Placing needs controlled force. The return move is predictable and repeatable.

If you try to handle all of that with a single control system, you are making tradeoffs everywhere. You end up with something mediocre across the board.

The straightforward answer is to use the right control method for each subtask. The hard part is doing that without turning every deployment into a manual segmentation exercise.

The tool

Skill segmentation takes the demonstrations you record and automatically divides each task into subtasks. For each one, it assigns the appropriate control method based on three factors: variability in the environment or the object, speed and speed requirements, and the complexity of the movement.

A high-variability subtask like a grasp gets a controller built to handle uncertainty. A fast, predictable transit gets something lightweight. A precise placement gets a force-aware controller. The return move gets the simplest thing that works.

The segmentation happens automatically from your demonstration data. You do not define subtask boundaries by hand. The system identifies them.

This is how you get a robot that is robust across the full task without over-engineering every step. It is also how you keep deployment timelines short. You record demonstrations, the segmentation runs, and you get a control stack that already reflects the structure of the task.

The tool is part of the same growing set of open dev tools we are releasing from our internal stack. If you are building robots and want to use it or contribute, reach out.