NVIDIA unveiled Halos for Robotics at the Automate 2026 conference in Chicago, a comprehensive robotics security system spanning chips, sensors, operating systems, and security certifications. Integrating over 18,600 engineering years of safety expertise and 7 million lines of verified code from NVIDIA’s autonomous driving domain, Halos provides a unified security architecture for autonomous robots. To date, 43 partners including Agility, Boston Dynamics, and Hesai have joined the ecosystem, with Agility already integrating Halos into its Digit robots and deploying them in facilities such as Amazon’s. The launch of Halos marks NVIDIA’s completion of the final piece in its end-to-end robotics stack, covering training, simulation, modeling, and security certification.

Article author, source: Quantum Bit

NVIDIA doesn't build robots, but it's helping embodied AI companies build better ones (doge)

Just now, at the Automate 2026 conference in Chicago, NVIDIA unveiled Halos for Robotics—

A full-stack robotics security system covering chips, sensors, operating systems, and security certifications.

The key feature of Halos is bringing NVIDIA’s over 18,600 engineer-years of safety expertise and 7 million lines of proven code from the autonomous driving domain to the robotics field, providing a unified safety architecture for autonomous robots.

With it, robotic companies no longer need to build everything from scratch—they can simply integrate and use it. More importantly, Halos’ core security framework has been open-sourced and made available to the industry.

One could say that if Tesla’s approach to embodied intelligence is like the iOS route—building its own robots and controlling its own security—then NVIDIA has chosen the Android route, opening its security platform to everyone.

Notably, numerous companies have already joined the Halos ecosystem as founding partners, including humanoid robotics companies Agility and Boston Dynamics, LiDAR manufacturer Hesai, and security robotics firm FORT Robotics, with the overall ecosystem now expanding to over 43 members.

Among them, Agility has been the first to adopt the "crab," integrating Halos into its Digit robots, which are now certified and operational in factories operated by Amazon, GXO, and Toyota.

The robot wearing a safety vest in the video is moving between factory conveyor belts, performing real-world tasks such as handling and logistics.

From chip to software, security spans three layers

What exactly is this new security system, Halos?

According to NVIDIA's official architecture, Halos can be divided into four layers, from bottom to top: Platform Security, Secure Operating System、Algorithm Security, and Ecosystem Security.

These four layers actually correspond to four dimensions of the same issue—

When robots operate in the real world, the four potential sources of errors are: hardware, software systems, model decisions, and external authentication and ecosystems.

First is platform security, ensuring the underlying hardware cannot be compromised.

NVIDIA has introduced the IGX Thor at this level, an AI computing platform designed for robotics and industrial applications.

It internally features a standalone "security island" with its own processor, I/O, power, and clock, physically isolated from the main computing system.

Even if the main AI system crashes, restarts, or operates abnormally, the safety island can still independently execute critical functions such as emergency braking.

It's similar to an aircraft's backup system, which can still take control when the primary system fails.

On the same layer is the Holoscan Sensor Bridge, designed to address another critical issue: latency and mismatch caused by sensor heterogeneity.

Robots are typically equipped with multiple devices such as LiDAR, depth cameras, IMUs, and torque sensors, but these devices come from different manufacturers and operate on different protocols.

If data needs to be processed in a层层 queue, the security window could be missed within tens of milliseconds.

The role of the Sensor Bridge is to unify the connection of all sensor data into a secure computing domain, enabling low-latency synchronized processing and achieving SIL 2 safety certification.

Layer 2: Secure operating system, addressing “whether the system itself can fail”

If the first layer ensures the hardware doesn’t fail, this layer ensures the system doesn’t collapse.

Halos OS runs on IGX Thor, with Halos Core as its underlying foundation, supporting two modes: pure Linux, or a hybrid Linux + QNX architecture.

In the latter, NVIDIA uses a hypervisor to split the system into two isolated domains: Linux handles AI computing and applications, while QNX manages safety-critical tasks. Both operate in complete isolation.

This means that even if there is an anomaly in the AI application layer, the security control logic remains unaffected. This layer acts as an additional “software isolation wall” outside the “hardware security island.”

Above this is the security application module, with the most typical example being the Outside-In Safety Blueprint.

The idea is: not only let the bot see the world on its own, but also introduce an external perspective.

For example, install cameras on the factory ceiling to monitor robot behavior from a third-party perspective using independent AI.

In a specific scenario, when an autonomous forklift operates inside a trailer, its onboard sensors may misjudge spatial boundaries, causing frequent emergency stops.

The Outside-In system can operate at higher efficiency when the environment is confirmed safe, and immediately take over to intervene if someone enters a hazardous area.

This capability is currently available to developers and is provided as open source.

Layer 3: Algorithmic security, addressing whether AI itself could make incorrect judgments

The first two layers ensure "system reliability," but the real risk from the bot comes from the higher layer—the model itself.

Whether it's a VLA (Vision-Language-Action model) or a VLM (Vision-Language model), their decisions may be incorrect.

For example, misclassifying a cardboard box as a person, or misclassifying a person as an obstacle. These errors are not system crashes, but rather “misunderstandings.”

The goal of this layer of algorithmic security is to evaluate and constrain the model’s behavioral safety in the physical world, ensuring that errors do not translate into dangerous actions.

Layer 4: Ecosystem Security, addressing "Who authenticates and who is responsible?"

At the top level is ecosystem security, responsible for turning the entire system into an industry standard.

NVIDIA established the Halos AI Systems Inspection Lab and obtained the world's first ISO/IEC 17020 accreditation in the field of physical AI. Certification bodies including TÜV Rheinland, TÜV SÜD, UL Solutions, SGS, exida, and CertX recognize its inspection results.

This means robotic companies can complete a preliminary inspection with NVIDIA before entering the formal certification process, significantly reducing time and cost.

In the past, this process was fragmented: sensors, controllers, and vision systems each had their own certifications and standards, requiring companies to assemble and recertify them independently.

Halos for the first time unifies the entire process—from chips and systems to models and certification—into a single integrated system.

Why does a bot need a "security system"?

Many of you likely had a similar question when you saw this news:

Why, after industrial robots have been reliably used for decades, is NVIDIA specifically launching a robotic safety system in 2026?

The reason is simple: embodied intelligent robots are now transitioning from laboratories to real-world industrial applications.

In the past, industrial robotic arms were fixed within workstations, their motion paths pre-programmed, and humans and machines were separated by barriers, with safety relying primarily on physical boundaries.

But now, a new generation of robots is entering factories, warehouses, and even offices, sharing the same space as humans.

Meanwhile, what drives them is no longer fixed rules, but embodied foundation models, distributed sensors, and real-time decision systems.

The change this brings is that bots are no longer "deterministic executors," but have become "agents with uncertainty."

Even in highly structured environments like factories, collaboration between different robots, material flow, changes in production line layout, and partial observability of the surrounding environment continuously introduce new risk variables.

This makes "security" no longer just a matter of mechanical isolation, but a system-level issue.

Regarding the necessity of security for robots entering factories, Agility CEO Peggy Johnson also stated:

For humanoid robots to create value at scale, safety must be built into the robots and validated at the system level. This is not optional—it is a prerequisite for humanoid robots to enter industrial workflows.

NVIDIA's Vice President of Robotics and Edge AI, Deepu Talla, takes this assessment a step further:

For robots to be scaled deployment in factories, warehouses, and logistics environments, the industry needs a unified security architecture.

In other words, the challenges the robotics industry faces today are similar to those autonomous driving faced over a decade ago—models are becoming increasingly intelligent, but what ultimately determines whether they can be deployed is often not the model itself, but safety.

And Halos is precisely NVIDIA's answer.

NVIDIA's full-stack system completes the final piece of the puzzle

In the end, NVIDIA’s full-stack robotics strategy has already taken shape.

If we break down this system, it can generally be divided into four layers: training, simulation, modeling, and inference.

Isaac Sim handles simulation training, enabling robots to learn how to interact with the world in a virtual environment;

GR00T provides foundational models that enable robots to understand commands, recognize their environment, and generate actions;

Cosmos builds a world model to predict how the physical world evolves under different actions;

Jetson Thor handles edge inference, enabling these capabilities to run directly on the robot itself.

From training to simulation, from models to deployment and inference, every layer of the entire technology pipeline is covered by NVIDIA products.

Now, Halos completes the final piece of the puzzle: security and access.

Once this process is complete, the bot is almost fully integrated into this technology stack.

Replacing any layer—especially the security and authentication system—would mean going through the verification process again, resulting in sunk time and costs.

Thus, the situation has become clear: NVIDIA does not manufacture robots, but it has established its interfaces at every level—from silicon to simulation, from models to safety certification.

This isn't just "helping you build a bot"; it's more like defining—

How are robots made?