Hi - I answer from the OpenSmartRoute documentation: routing, the API, plans and quotas, self-hosting. Ask away, or open a support ticket if you need a person.
Grounded in the docs - follow a source before acting on it.
Safeworld raises $12 million for AI robot safety standards - OpenSmartRoute
Safeworld launches with major funding from top investors
Safeworld is a new company focused on making generative AI robots safe. The team launched the business after working on safety problems for years. Dr. Ding Zhao leads the Safe AI lab at Carnegie Mellon University. He has spent almost his entire career studying these issues. Now he founded Safeworld with two other experts.
The company just raised a seed round of more than $12 million. This is a significant amount of money for a startup in stealth mode. Shine Capital led this investment round as the primary backer. a16z Speedrun was also a major investor in the deal. Box Group provided additional funding to support their early growth.
Carnegie Mellon University Endowment invested alongside other partners. Innovation Endeavors and SV Angel joined the group of investors. These backers believe safety is critical before robots enter homes or factories. Jonathan Lai from a16z Speedrun emphasized the urgency of this work. He told TechCrunch that timing is everything for building industry standards.
"The time to build an industry safety standard is now," Lai said. "By the time you have robots in households colliding with kids, that's way too late." This quote highlights the risk of waiting until accidents happen. Safeworld aims to prevent these incidents before they occur. They want to set rules while robots are still being designed.
The founders hope to become the first profitable company in this specific field. Dr. Zhao believes people will pay for safety guarantees. If anyone wants to deploy robots, they need reliable safety checks. This creates a potential revenue stream for Safeworld's services. The team is confident they are solving a vital problem for the industry.
The core problem of unpredictable generative AI in robotics
The biggest trend in robotics is handing control over to generative AI models. However, this architecture introduces a major unpredictability issue. Traditional algorithms work differently than these new probabilistic systems. You cannot be sure how a generative model will react in every situation.
Dr. Zhao describes the safety challenge as having two difficult parts. The first part involves advanced generative AI probabilistic evals. This means figuring out how to underwrite the risk of a probabilistic system. The second part is building trust with people who will use the robots. You need both risk assessment and trust to deploy a robot safely.
OpenAI plans to dump hundreds of AI-solved math problems on GitHub without publishing papers. Mathematicians want formal verification and proper credit before accepting the results.
Apple and Google are building devices that process data without saving it as a traditional recording.
Generative AI models do not follow fixed rules like old math equations. They make decisions based on patterns they learned during training. This makes their behavior hard to predict in complex environments. Traditional verification methods rely on formal proofs and mathematical certainty. These methods fail when the system involves randomness or open-ended reasoning.
Robot makers face a dilemma when adopting these new technologies. They want the flexibility of generative AI but fear the lack of predictability. How can you be sure your brand new humanoid will be safe? The answer lies in rigorous testing and simulation before real-world deployment.
How Safeworld builds digital simulations for safety testing
Safeworld specializes in evaluating robotic control systems within digital simulations. These simulations are populated with realistic human models to test robot behavior. It is similar to the work companies like Tesla or Wayve do for their vehicles. They ensure cars respond appropriately to surprising incidents on the road.
But robots face more difficult challenges than self-driving cars. They operate in unstructured environments that change constantly. Each facility they enter has different safety standards and physical layouts. Safeworld builds a digital version of these specific corners and scenarios.
They use models like Genesis or MuJoCo for their simulations. These tools create physics-based worlds where robots can interact with humans. The team inserts the simulation of the robot being evaluated into this world. They drive the robot using its real software code, not a simplified model.
Then they run thousands of scenarios where human models encounter the robot. This process is harder than it seems because people are unpredictable. Tripping and falling are common events that require extensive testing. Otherwise, you would have to trip and fall for the robot constantly. That is like a hard thing to be doing all the time in reality.
Why traditional math verification fails for modern robots
Many robotic systems cannot be formally proven safe using standard mathematics. Vishal Dugar, the CTO of Gritt Robotics, explains this limitation clearly. His company develops AI brains for robots that help workers install solar panels. He notes that it is very hard to formally prove safety by doing some math.
Writing some equations and saying "yeah, the system is verified to be safe" does not work here. Dugar says the difficulty with most of their systems lies in this lack of formal proof. It necessarily has to be done empirically through real-world or simulated testing. Mathematical verification cannot account for every possible human interaction or edge case.
The robots operate alongside human workers in industrial settings. Ensuring that a robotic arm does not hit them is obviously top of mind for engineers. To verify this in practice requires considering all kinds of potential scenarios. You must test how the robot reacts to kneeling, standing, tripping, crouching, or running humans.
Humans have many kinds of appearances and body configurations. Their bodies can be in different positions depending on what they are doing. They could be wearing different clothes, varying in size, shape, height, or skin color. You have to respond to all these behaviors that humans could potentially exhibit on these sites. The variety in human appearance adds another layer of complexity to safety verification.
Real-world challenges like blind corners and human variety
One of the most common areas for danger is if there is a blind corner in a factory. Kyle Wong from Safeworld points out this specific scenario as a frequent testing ground. What is the speed or what is the stopping distance that you need? You must make sure that this robot will not collide with a particular human.
If a human is carrying boxes, for example, will the robot detect the human or not? This question highlights the difficulty of recognizing objects in cluttered or obstructed views. The robot must understand context beyond simple object detection. It needs to anticipate potential hazards that humans create in their environment.
The founders believe robot-makers will want a third-party to validate their work. They may share information about safety cases between competitors for transparency. A lot of people are underestimating how hard some of these edge cases are going to be to solve. It is not the robot in the vacuum, in the demo, that we are worried about.
It is the robot that is deployed at scale, with people who potentially never operated a robot before. These users bring their own unique behaviors and expectations into the shared space. The system must adapt to this human variety without causing harm or confusion.
Why it matters for deploying safe agents at scale
The time to build an industry safety standard is now while robots are being designed. Waiting until robots are in households colliding with kids causes way too many safety incidents. Jonathan Lai from a16z Speedrun made this point clearly during the funding announcement. The cost of fixing accidents after they happen far outweighs the cost of prevention.
Deploying safe agents at scale requires more than just good hardware or software. It requires a robust framework for evaluating and certifying robot behavior. Companies like Gritt Robotics are partnering with Safeworld to develop these safety simulations. They need external validation to prove their systems do not hit human workers.
The team is still figuring out the best model for its product. They are deciding between a platform for external users or a services-based approach. Regardless of the business model, the core mission remains focused on safety. If anyone wants to deploy, they need to pay us to handle the situation. This revenue model suggests high demand for their safety certification services.
What to do when running your own robotic systems
Engineers and managers should check if their robotic systems use generative AI models. You need to understand how these probabilistic systems differ from traditional algorithms. Look at the documentation to see if formal math verification is possible or missing.
Consider partnering with third-party safety organizations for validation. Companies like Safeworld offer specialized testing in digital simulations. They can help you identify edge cases that your internal teams might miss. Ask them about their simulation tools and whether they support your specific robot models.
Review the scenarios where your robots interact with humans in unstructured environments. Think about blind corners, varying human appearances, and unexpected behaviors like tripping. Run thousands of these scenarios in a simulator before deploying to real facilities. This step is harder than it seems but essential for safety.
Monitor your robots closely during initial deployments to catch any emergent unsafe behaviors. You cannot trip and fall for the robot all the time in the real world. Real-world testing must supplement simulation results to ensure comprehensive coverage. Stay updated on industry standards as they evolve rapidly with new AI models.
