Programming robots | LIVE141

Chip Huyen and John Maeda discuss practical AI systems for robotics, focusing on what it takes to program real robots safely and consistently. They cover the Vision-Language-Action (VLA) approach, challenges in collecting robot action data, and the idea of interfacing with multiple robots through a shared API.

Overview

This Microsoft Build 2026 session looks at how AI is moving from purely digital applications into the physical world, and what changes when “the app” is a robot.

Key themes covered in the session include:

• How to program robots to perform tasks reliably in real environments.
• Using a consistent API surface to interface with multiple robot platforms.
• The Vision-Language-Action (VLA) approach and why it matters for robotics.
• Practical constraints around collecting and using robot action data.
• Safety-by-design examples for robots operating around humans (for example, intentionally weak grip strength and maintaining safe distances).

Session structure (from the published chapters)

Overview: practical AI systems and robotics

• Framing: AI capabilities are rapidly evolving, and robotics adds real-world constraints (safety, data collection, physical interaction).

Unitary and robotics context

• Introduction to Unitary as a profitable robotics company.

Robot motion via pre-captured movement demos

• Discussion of improving robot motion using pre-captured movement demonstrations.

Vision-Language-Action (VLA)

• Explanation of the VLA approach.

Collecting robot action data

• Challenges in collecting the data needed to train or improve robot behaviors.

Rapidly evolving AI primitives

• How fast-moving AI building blocks affect developers building real systems.

Safety design examples

• Example safety design from a 1X robot: intentionally weak grip for safety.
• Design rules for maintaining safe distance between robots and humans.