Abstrae is building robots that can be easily debugged. Unlike current approaches like VLA models, the goal is to create a framework that allows users to visualize action plans and verify their coherence. If they are not coherent, the teacher can adjust the plan and add missing information that the system may have overlooked.

This may sound like a less powerful system because humans must design action plans for every new task, but it enables robots that are not mere automatons—and, most importantly, that provide understandable feedback. Over time, with community contributions, more action plans and feature models will be developed, so users will not always need to design and debug them.

Current robotics companies focus on general-purpose robots with pre-trained, non-reprogrammable models. To adapt one to a new task, users must wait for the company to train the model for that specific task.

The core challenge lies in converting high-level action plans—represented as graphs—into lower-level actions for joint actuation. In the first few months, the community will help build models that the robot can integrate. These are not models the community must train, but rather directions and concepts guiding the robot to succeed. This involves specifying relevant features and their connections to the environment. For example, to place hangers on a bar, users can define points of contact and the required orientation.

Ideally, users shouldn't program. They should act as instructors, giving tips to ensure task success. Any necessary programming should be handled by Abstrae, with the framework making it easy to explain tasks to the robot.

Robots must also adapt when tasks fail. Users may reach the limits of the robot's knowledge, especially in cases requiring precise control, like peeling a vegetable, where no clear concepts exist for micro-adjustments. This will occur frequently and is a critical challenge. Abstrae is actively developing a method that enables the robot to learn these micro-adjustments after just a few trials, with human guidance.