Nvidia is working on new safety and reliability systems aimed at making humanoid robots capable of safely operating in close proximity to people in real-world environments like warehouses, factories and eventually homes, Bloomberg reports.
The central challenge the company is addressing is that humanoid robots will need to act autonomously in dynamic, unpredictable settings where humans are constantly moving around them.
That requires much more than basic object detection. It involves real-time perception, prediction of human movement and the ability to instantly adjust actions to avoid collisions or unsafe behavior.
To solve this, Nvidia is building what it describes as a “physical AI” safety stack. This combines onboard robot computing with external sensing systems and software layers that continuously evaluate a robot’s surroundings. The idea is to create redundancy so that even if one system misreads a situation, another can intervene or override decisions to maintain safety.
A major component of this effort is a robotics safety framework called “Halos.” It is designed to standardize how robots are tested, validated and monitored before and during deployment.
Halos draws from Nvidia’s experience in autonomous driving safety systems, adapting similar concepts like sensor fusion, real-time constraints, fail-safes and simulation-based testing to humanoid robotics.
The system also emphasizes continuous monitoring and feedback, meaning robots are not just pre-programmed and released, but are constantly assessed and improved through data collected during operation. This is intended to reduce unexpected behaviors that could lead to harm in human-robot interactions.
