Safety Instinct for Physical AI: A CBF-Based Safety Cerebellum for Quadruped Robots

• Role: independent research, conceived and completed independently
• Period: Feb 2026 – Apr 2026
• Affiliation: Inchor (SaferDrive AI), during the Simulation Development Engineer internship
• Status: completed; an internal exploration

Overview

An internal exploration of a “safety instinct” for Physical AI: a safety cerebellum for a Boston Dynamics Spot robot, built in NVIDIA Isaac Sim (Omniverse).

Methodology

A reinforcement learning (RL) waypoint-following controller was trained for Spot, and a Control Barrier Function (CBF) safety filter was layered beneath it: the learned policy handles navigation while the CBF layer enforces safety.

Contribution

Spot maintains a dynamic safety boundary around moving obstacles while traveling between waypoints.

Tools

NVIDIA Isaac Sim (Omniverse), reinforcement learning, Control Barrier Functions