The Case for a Safety-First Foundation in Embodied AI

Many companies fall into a familiar trap when building robots: They first develop their technology, only to treat safety as an afterthought. This approach is risky and inefficient, and when it comes to embodied artificial intelligence (AI) — where AI operates in and interacts with the physical world — the stakes are even higher.

As embodied AI continues to evolve and integrate more deeply into the world around us, it’s essential to remember that safety must be the foundation on which these systems are built.

The Dangers of Treating Safety as an Afterthought

Guardrails for large language models (LLMs) are already a widely recognized necessity in the software world. However, they’re even more critical in embodied AI, where failures can lead to physical harm rather than mere inconvenience or user frustration.

Most companies recognize the risks of neglecting safety or operating without proper guardrails. Yet safety is still too often viewed as a barrier to innovation, with engineers frequently prioritizing functionality and treating safety as a “to-do” for later. In reality, deferring safety considerations can lead to costly inefficiencies and setbacks.

Early-stage robotics companies, for example, frequently focus on demonstrating capabilities to investors. Although this approach may secure short-term funding, it generates technical debt that grows increasingly expensive to resolve as development progresses.

Retrofitting safety mechanisms into systems not initially designed for them is exponentially more complex than building them in from the start. This challenge is compounded when there’s a mismatch between development skills and safety expertise.

Building Safety into the Foundation

Starting with safety from the onset of a project offers a more efficient and scalable path for developing embodied AI systems. One powerful tool in enabling safety-first design is control barrier functions (CBFs). These mathematical tools provide a formal approach to ensuring that a system will remain within safe operating parameters while allowing maximum freedom to achieve its objectives. 

Think of CBFs as an invisible wall or a force field that gently redirects a robot away from unsafe situations with minimal disruption to its primary task. Unlike rigid safety measures that can shut down operations entirely, these functions apply the minimum correction necessary to maintain safety. 

We saw this “force field” in action in 2024, when the U.S. Air Force Test Pilot School equipped the X-62A VISTA (Variable In-flight Simulation Test Aircraft) with 3Laws’ Guardrails technology. The system enforced preset flight constraints, such as geofencing, altitude ceilings, and G-force limits, automatically preventing pilots from exceeding those parameters. This kind of built-in safety will be critical to advancing autonomous military aircraft from test environments to real-world deployment.

Designing for Safety, Unlocking Performance

Robots with rigid safety measures often resort to full-stop behaviors when encountering unexpected scenarios. These abrupt halts don’t just slow operations; they can also damage equipment and create production bottlenecks. 

Another way to look at CBFs is as bowling bumpers for innovation. When integrated early into the development phase, they give engineers the confidence to experiment boldly without fearing catastrophic failures. With safety mathematically justified and built into the system architecture, engineers needn’t rely solely on exhaustive testing of their autonomy layer to ensure the robot will respond appropriately in every scenario.

While safety has historically been seen as a barrier to performance, a safety-first approach can accelerate development. This marks a decisive shift in a field long constrained by safety concerns. The result is paradoxical but proven: Innovation moves faster when safety is the foundation.

Case Study: Manipulator Autonomy Development

To provide more context, imagine you’re developing the autonomy layer for a robotic arm (manipulator) that works in close proximity to humans within a complex environment, such as in a warehouse or assembly line. You have three main design approaches to consider:

• Option 1: No safety system. Operating without a safety system exposes you to significant risk and liability. In this scenario, the autonomy stack must be trusted and tested extensively in a safe environment to test in the real environment. This results in a slow product rollout and slows down updates and changes. Plus, since real-world environments are inherently unpredictable, it’s impossible to anticipate and test for every scenario, leaving a constant margin of risk.
• Option 2: A traditional safety system. This approach reduces the pressure on the autonomy stack to be inherently safe, but it introduces significant challenges during testing. If a safety incident is triggered, the system immediately e-stops, halting the test. Because the safety system operates independently of the autonomy stack, you don’t gain any meaningful insight into what caused the failure, making it difficult to diagnose issues or improve system behavior.
• Option 3: An intelligent, dynamic safety layer. With this approach, you can test your autonomy without worrying about safety risks or the safety system disrupting your testing or operations. When the autonomy stack makes a mistake (which is inevitable), the system keeps operating and the dynamic safety layer will give you insights as to what the autonomy stack tried to do and why it was unsafe. 

With option 3, developing autonomy becomes significantly easier. You can test faster without concern while gaining valuable insights into both the performance and safety of the motion planning.

Safety as a Catalyst for Innovation 

Safety shouldn’t be something we bolt onto embodied AI. Rather, it should be used as a foundation to build on. By embedding collision avoidance based on control barrier functions into the system, teams can test new algorithms without risking equipment or human safety. It enables faster iteration, more ambitious testing, and ultimately, quicker delivery of a reliable product.

It also doesn’t have to slow you down. When safety is prioritized from day one and properly implemented, it’s possible to move fast — without breaking things.