The Case for Nvidia’s AV Safety Force Field (SFF)

Maggie Zhang
A Study in AI Ethics
2 min readFeb 13, 2020
Image from Nvidia’s Safety Force Field promotional website: https://www.nvidia.com/en-us/self-driving-cars/safety-force-field/

In 2017, Mobileye published a framework for autonomous vehicle safety which it called Responsibility-Sensitive Safety (RSS) and called on other industry players to adopt the framework as a starting point for an industry standard. In 2019, Nvidia published its own version of an AV safety framework called Safety Force Field (SFF). Immediate backlash from Intel ensued, with Mobileye co-founder and CEO Amnon Shashua calling out Nvidia on how similar SFF was to RSS and causing unnecessary confusion in the AV safety space.

Now that the backlash has subsided, I’d like to make a case for Nvidia’s SFF AV safety framework. While much of the framework builds on top of RSS, it is helping to normalize the notion of a mathematical model for AV safety and spread awareness of safety standards to OEMs and AV fleet production companies as well as the public.

Reason 1: SFF builds on top of the existing RSS model

Nvidia did not create an entirely new AV safety framework, but instead chose to modify the language around safety, which is an ever-evolving definition that can be mathematically defined in many ways. According to David Nister, VP of Autonomous Driving Software at Nvidia, SFF is built on a core principle of collision avoidance as opposed to the numerous rules and exceptions outlined in RSS. Nister says that SFF has a more constrained set of priorities, such as not causing a crash and mitigating unsafe conditions. SFF also operates independently of the main AV control stack which provides a degree of redundancy in case the main control stack is compromised.

Reason 2: RSS and SFF are not mutually exclusive

The culture right now surrounding safety in AVs should not be about “who created the first model” but rather, “how can we best collaborate with the public and other companies towards a comprehensive safety model”. Since both RSS and SFF are open models, anyone could combine elements from both to build an even better model.

Reason 3: Nvidia’s publishing of SFF boosts trust in self-driving vehicles

The publishing of RSS and SFF does not necessarily or should cause confusion — in fact, it should give OEMs and fleets building AVs more options to think about. Increasing conversation around safety will boost trust in self-driving vehicles. Mobileye should not feel attacked but should view Nvidia’s publication as a step into bringing a wider adoption of safety frameworks to companies.

Ultimately, it’s cross-collaboration but more importantly iterating on existing safety frameworks that will drive adoption of autonomous vehicles. RSS and SFF should not be treated as mutually exclusive safety frameworks, but should be viewed as the first steps towards building trust between AV companies and the public.

--

--