A robot can pick up a box, navigate a warehouse, and work alongside human workers. But what happens when it misreads an object, encounters an obstacle, or makes an unexpected decision?
In July this year, a viral video showed a Chinese humanoid robot in Saratov, Russia, unable to recognise the gesture of a handshake and taking a fighting stance before attempting to kick the customer.
While that video may or may not have been staged, AI-powered robots are moving toward industrial deployment, and that presents a challenge beyond building capable machines. They must prove these systems can operate safely in the real world.
NVIDIA’s Halos for robotics aims to address this challenge through safety across the computing stack. Amit Goel, NVIDIA’s Head of Robotics Ecosystem and Edge AI Products, tells AIM that conventional safeguards cannot support every task that modern robots must perform.
Addverb, which develops industrial robotics systems, sees another hurdle. Its testing has revealed issues ranging from incorrect depth calculations to minor collisions. These experiences highlight why robots must continuously validate their actions, rather than simply execute instructions.
The stakes are rising as physical AI models become production-ready. Robots must now navigate environments where people, machines, and tasks can change constantly.
When Stopping is No Longer Enough
Industrial robotics has long relied on predictable environments and predefined safety boundaries. Goel says conventional systems often stop when they detect an object nearby, without distinguishing between people and obstacles.
“The second situation was an autonomous mobile robot that is moving around carrying stuff,” Goel says. “And while it was not inside a cage, it was existing around people.”
Such systems can work in specific industrial applications. However, frequent stops and restricted operating areas limit their usefulness across different tasks.
Goel believes the old way of using laser curtains, or simply bolting on safety as a sensor, is not sufficient. Robots powered by foundation models need to interpret situations before deciding how to act. A robot handing an object to a worker must approach them. The same proximity could present a risk during another task.
“The context and the reasoning are a critical part of your safety guarantees or safety function that you’re writing,” Goel says.
NVIDIA developed Halos as a reference platform for building safety into physical AI systems. The stack covers computing hardware, system design, operating systems, and safety applications.
The company also developed Holoscan Sensor Bridge to bring sensor data into computing systems with low latency. Its architecture separates safety logic from the robot’s main functions, allowing both to operate on the same computer.
External Cameras
One part of NVIDIA’s approach is Outside-In Safety. It uses cameras installed around a facility to monitor interactions between robots, people, and machinery. A robot might struggle to see around a blind corner in a warehouse. Slowing down before every turn improves safety, but can reduce operational throughput.
NVIDIA’s reference application uses external camera feeds to help identify potential hazards. It can then send commands to robots when their movements could create unsafe situations.
Goel says developers can adapt this approach across different environments. These include factories, warehouses, hospitals, and retail spaces, each with its own safety requirements.
Simulation also plays a role in testing interactions between multiple AI systems. Goel explains that developers need to understand how several models behave when operating together inside a facility.
The company also runs an inspection lab for developers building on its safety stack. The lab reviews technical documentation and provides feedback before developers approach certification bodies.
Goel clarifies that the lab does not certify products itself. Instead, its inspections aim to help developers prepare for formal certification. “AI and safety is a very new thing,” he says.
He added that standards bodies are still evolving their approaches to these systems. The diversity of robots, sensors, and applications makes it difficult to develop common safety solutions.
Between Successful Pilot and Production
For Addverb, the challenge often appears during testing. The company reported depth miscalculations when stock-keeping units overlap, gripper over-rotation during recovery, and minor collisions during retraction.
These problems can emerge even when a robot performs its intended task. Rajesh Kumar, Head of Advanced Robotics at Addverb, says its systems need redundant depth sensing, confidence scoring, and in-motion checks to catch errors.
When a robot encounters an unexpected obstacle or unfamiliar task, its system combines sensor data, confidence levels, and predefined controls. It can slow down, scan again, re-plan its movement, or stop.
If the system cannot proceed safely, it can hand the task to a human. Speed limits, torque limits, emergency stops, and geofencing also operate independently of the learning system.
Kumar believes it validates robots through supervised proof-of-concept deployments in relevant operating environments. These tests assess task performance, obstacle handling, stopping behaviour, and compliance with safety boundaries.
However, he says that no single threshold applies to every deployment. Requirements depend on the task and the environment in which the robot operates.
“The biggest gap is consistency in unpredictable environments,” Kumar tells AIM.
A successful pilot proves that a robot can perform a task, but production requires reliable performance across variations in objects, surroundings and operating conditions.
That distinction becomes more important as robots gain the ability to perform a wider range of tasks. Developers must also investigate failures across the entire system, rather than automatically blaming the AI model.
NVIDIA is addressing another risk through OpenShell, an open project designed to sandbox AI models running on robots. Developers can define policies that prevent models from exceeding limits, such as maximum torque.
The company has also described an out-of-band monitoring system designed to detect and block unsafe activity. These safeguards aim to constrain what an AI model can do, even when its decisions exceed expected boundaries.
Safety remains one of several barriers to wider humanoid deployment, alongside computing efficiency and the maturity of AI models. Goel acknowledges that the industry still has work ahead.
For now, the challenge is not simply making robots more intelligent. It is ensuring their growing capabilities stay within limits people can trust.
