
Gemini Robotics On-Device 2 Makes Physical AI a Data Problem
Google DeepMind’s on-device robotics work shifts attention from impressive demonstrations to the data, latency, and safety constraints that make robots useful.
Gemini Robotics On-Device 2 Makes Physical AI a Data Problem
A robot that waits for a round trip to the cloud is not autonomous in the way a person experiences autonomy. It may be connected, but every delay becomes a physical event: a hand pauses above a shelf, a gripper misses a moving object, or a safety controller has to take over. Google DeepMind’s Gemini Robotics On-Device 2 model card puts the important question in a more practical place. The future of physical AI may depend less on a bigger model than on a model that can learn from local data and act inside a tight latency budget.
On-device is a deployment constraint, not a slogan
Google DeepMind’s model-card page describes Gemini Robotics On-Device 2 as an embodied model intended to run locally on robot hardware: https://deepmind.google/models/model-cards/gemini-robotics-on-device-2/. The phrase “on-device” changes the design target. Memory, thermal limits, sensor timing, and intermittent connectivity matter as much as language capability. A cloud model can spend more compute deciding what a scene means; a robot must decide whether the delay itself creates risk.
Google DeepMind’s model-card page describes Gemini Robotics On-Device 2 as an embodied model intended to run locally on robot hardware: https://deepmind.google/models/model-cards/gemini-robotics-on-device-2/. The phrase “on-device” changes the design target. Memory, thermal limits, sensor timing, and intermittent connectivity matter as much as language capability. A cloud model can spend more compute deciding what a scene means; a robot must decide whether the delay itself creates risk. The first reporting question is not whether the idea sounds ambitious, but which dependency could make it fail in an ordinary production environment. The practical implication is that readers should separate what the source directly establishes from what follows as analysis. That discipline matters for gemini robotics on-device 2 makes physical ai a data problem: a plausible mechanism is not the same thing as observed performance, and a vendor, research group, or forecaster should be credited for its claim rather than treated as a neutral narrator.
Google DeepMind’s model-card page describes Gemini Robotics On-Device 2 as an embodied model intended to run locally on robot hardware: https://deepmind.google/models/model-cards/gemini-robotics-on-device-2/. The phrase “on-device” changes the design target. Memory, thermal limits, sensor timing, and intermittent connectivity matter as much as language capability. A cloud model can spend more compute deciding what a scene means; a robot must decide whether the delay itself creates risk. A second question concerns incentives. The organization making the announcement benefits from a favorable interpretation, while customers, workers, or operators carry different risks. Good analysis keeps those positions visible and asks what independent measurement would change the conclusion. The practical implication is that readers should separate what the source directly establishes from what follows as analysis. That discipline matters for gemini robotics on-device 2 makes physical ai a data problem: a plausible mechanism is not the same thing as observed performance, and a vendor, research group, or forecaster should be credited for its claim rather than treated as a neutral narrator.
The model sees a world that keeps moving
Images and language are forgiving compared with contact. A robot must account for occlusion, friction, lighting, joint limits, calibration drift, and objects that do not behave like training examples. DeepMind’s robotics research explains the broader embodied-reasoning direction: https://deepmind.google/blog/gemini-robotics-er-1-6/. The model is useful only when perception and action remain synchronized. A correct description arriving too late is still a failed control decision.
Images and language are forgiving compared with contact. A robot must account for occlusion, friction, lighting, joint limits, calibration drift, and objects that do not behave like training examples. DeepMind’s robotics research explains the broader embodied-reasoning direction: https://deepmind.google/blog/gemini-robotics-er-1-6/. The model is useful only when perception and action remain synchronized. A correct description arriving too late is still a failed control decision. The first reporting question is not whether the idea sounds ambitious, but which dependency could make it fail in an ordinary production environment. The practical implication is that readers should separate what the source directly establishes from what follows as analysis. That discipline matters for gemini robotics on-device 2 makes physical ai a data problem: a plausible mechanism is not the same thing as observed performance, and a vendor, research group, or forecaster should be credited for its claim rather than treated as a neutral narrator.
Images and language are forgiving compared with contact. A robot must account for occlusion, friction, lighting, joint limits, calibration drift, and objects that do not behave like training examples. DeepMind’s robotics research explains the broader embodied-reasoning direction: https://deepmind.google/blog/gemini-robotics-er-1-6/. The model is useful only when perception and action remain synchronized. A correct description arriving too late is still a failed control decision. A second question concerns incentives. The organization making the announcement benefits from a favorable interpretation, while customers, workers, or operators carry different risks. Good analysis keeps those positions visible and asks what independent measurement would change the conclusion. The practical implication is that readers should separate what the source directly establishes from what follows as analysis. That discipline matters for gemini robotics on-device 2 makes physical ai a data problem: a plausible mechanism is not the same thing as observed performance, and a vendor, research group, or forecaster should be credited for its claim rather than treated as a neutral narrator.
Data collection becomes the strategic bottleneck
Robotics data is expensive because it is attached to hardware, time, and physical risk. A web-scale corpus cannot fully describe how a particular gripper behaves on a particular surface. The research paper on scaling learning across robot types shows why diversity and transfer are central problems: https://deepmind.google/blog/scaling-up-learning-across-many-different-robot-types/. On-device inference helps deployment, but it does not remove the need for carefully labeled demonstrations and recovery traces.
Robotics data is expensive because it is attached to hardware, time, and physical risk. A web-scale corpus cannot fully describe how a particular gripper behaves on a particular surface. The research paper on scaling learning across robot types shows why diversity and transfer are central problems: https://deepmind.google/blog/scaling-up-learning-across-many-different-robot-types/. On-device inference helps deployment, but it does not remove the need for carefully labeled demonstrations and recovery traces. The first reporting question is not whether the idea sounds ambitious, but which dependency could make it fail in an ordinary production environment. The practical implication is that readers should separate what the source directly establishes from what follows as analysis. That discipline matters for gemini robotics on-device 2 makes physical ai a data problem: a plausible mechanism is not the same thing as observed performance, and a vendor, research group, or forecaster should be credited for its claim rather than treated as a neutral narrator.
Robotics data is expensive because it is attached to hardware, time, and physical risk. A web-scale corpus cannot fully describe how a particular gripper behaves on a particular surface. The research paper on scaling learning across robot types shows why diversity and transfer are central problems: https://deepmind.google/blog/scaling-up-learning-across-many-different-robot-types/. On-device inference helps deployment, but it does not remove the need for carefully labeled demonstrations and recovery traces. A second question concerns incentives. The organization making the announcement benefits from a favorable interpretation, while customers, workers, or operators carry different risks. Good analysis keeps those positions visible and asks what independent measurement would change the conclusion. The practical implication is that readers should separate what the source directly establishes from what follows as analysis. That discipline matters for gemini robotics on-device 2 makes physical ai a data problem: a plausible mechanism is not the same thing as observed performance, and a vendor, research group, or forecaster should be credited for its claim rather than treated as a neutral narrator.
Generalization needs a failure vocabulary
A benchmark score can hide the failures that matter in a warehouse or home. Engineers need to record near misses, ambiguous instructions, dropped objects, unexpected people, and cases where the robot should refuse. The Robotics Industry Association’s safety guidance provides a useful operational frame: https://www.automate.org/robotics/safety. A model that knows how to say “I cannot safely complete this action” may be more valuable than one that attempts a wider set of tasks.
A benchmark score can hide the failures that matter in a warehouse or home. Engineers need to record near misses, ambiguous instructions, dropped objects, unexpected people, and cases where the robot should refuse. The Robotics Industry Association’s safety guidance provides a useful operational frame: https://www.automate.org/robotics/safety. A model that knows how to say “I cannot safely complete this action” may be more valuable than one that attempts a wider set of tasks. The first reporting question is not whether the idea sounds ambitious, but which dependency could make it fail in an ordinary production environment. The practical implication is that readers should separate what the source directly establishes from what follows as analysis. That discipline matters for gemini robotics on-device 2 makes physical ai a data problem: a plausible mechanism is not the same thing as observed performance, and a vendor, research group, or forecaster should be credited for its claim rather than treated as a neutral narrator.
A benchmark score can hide the failures that matter in a warehouse or home. Engineers need to record near misses, ambiguous instructions, dropped objects, unexpected people, and cases where the robot should refuse. The Robotics Industry Association’s safety guidance provides a useful operational frame: https://www.automate.org/robotics/safety. A model that knows how to say “I cannot safely complete this action” may be more valuable than one that attempts a wider set of tasks. A second question concerns incentives. The organization making the announcement benefits from a favorable interpretation, while customers, workers, or operators carry different risks. Good analysis keeps those positions visible and asks what independent measurement would change the conclusion. The practical implication is that readers should separate what the source directly establishes from what follows as analysis. That discipline matters for gemini robotics on-device 2 makes physical ai a data problem: a plausible mechanism is not the same thing as observed performance, and a vendor, research group, or forecaster should be credited for its claim rather than treated as a neutral narrator.
Latency is an architectural budget
Local execution turns latency into a budget that can be allocated across sensing, perception, planning, control, and actuation. NVIDIA’s Isaac documentation describes the simulation and deployment layers that robotics teams use to manage this pipeline: https://developer.nvidia.com/isaac. A model can be impressive in isolation and still be unsuitable if preprocessing consumes the time required for stable control. The measurement must be end to end, from sensor frame to physical action.
Local execution turns latency into a budget that can be allocated across sensing, perception, planning, control, and actuation. NVIDIA’s Isaac documentation describes the simulation and deployment layers that robotics teams use to manage this pipeline: https://developer.nvidia.com/isaac. A model can be impressive in isolation and still be unsuitable if preprocessing consumes the time required for stable control. The measurement must be end to end, from sensor frame to physical action. The first reporting question is not whether the idea sounds ambitious, but which dependency could make it fail in an ordinary production environment. The practical implication is that readers should separate what the source directly establishes from what follows as analysis. That discipline matters for gemini robotics on-device 2 makes physical ai a data problem: a plausible mechanism is not the same thing as observed performance, and a vendor, research group, or forecaster should be credited for its claim rather than treated as a neutral narrator.
Local execution turns latency into a budget that can be allocated across sensing, perception, planning, control, and actuation. NVIDIA’s Isaac documentation describes the simulation and deployment layers that robotics teams use to manage this pipeline: https://developer.nvidia.com/isaac. A model can be impressive in isolation and still be unsuitable if preprocessing consumes the time required for stable control. The measurement must be end to end, from sensor frame to physical action. A second question concerns incentives. The organization making the announcement benefits from a favorable interpretation, while customers, workers, or operators carry different risks. Good analysis keeps those positions visible and asks what independent measurement would change the conclusion. The practical implication is that readers should separate what the source directly establishes from what follows as analysis. That discipline matters for gemini robotics on-device 2 makes physical ai a data problem: a plausible mechanism is not the same thing as observed performance, and a vendor, research group, or forecaster should be credited for its claim rather than treated as a neutral narrator.
Simulation helps, but reality collects the bill
Simulation can multiply training examples and expose rare conditions without damaging equipment. It also introduces a gap between rendered surfaces and real contact. NVIDIA’s Cosmos and Isaac research illustrates the industry’s effort to make simulated environments more useful for physical AI: https://developer.nvidia.com/isaac/sim. The honest standard is not whether a policy succeeds in simulation; it is how much real-world calibration is required before people can trust it.
Simulation can multiply training examples and expose rare conditions without damaging equipment. It also introduces a gap between rendered surfaces and real contact. NVIDIA’s Cosmos and Isaac research illustrates the industry’s effort to make simulated environments more useful for physical AI: https://developer.nvidia.com/isaac/sim. The honest standard is not whether a policy succeeds in simulation; it is how much real-world calibration is required before people can trust it. The first reporting question is not whether the idea sounds ambitious, but which dependency could make it fail in an ordinary production environment. The practical implication is that readers should separate what the source directly establishes from what follows as analysis. That discipline matters for gemini robotics on-device 2 makes physical ai a data problem: a plausible mechanism is not the same thing as observed performance, and a vendor, research group, or forecaster should be credited for its claim rather than treated as a neutral narrator.
Simulation can multiply training examples and expose rare conditions without damaging equipment. It also introduces a gap between rendered surfaces and real contact. NVIDIA’s Cosmos and Isaac research illustrates the industry’s effort to make simulated environments more useful for physical AI: https://developer.nvidia.com/isaac/sim. The honest standard is not whether a policy succeeds in simulation; it is how much real-world calibration is required before people can trust it. A second question concerns incentives. The organization making the announcement benefits from a favorable interpretation, while customers, workers, or operators carry different risks. Good analysis keeps those positions visible and asks what independent measurement would change the conclusion. The practical implication is that readers should separate what the source directly establishes from what follows as analysis. That discipline matters for gemini robotics on-device 2 makes physical ai a data problem: a plausible mechanism is not the same thing as observed performance, and a vendor, research group, or forecaster should be credited for its claim rather than treated as a neutral narrator.
flowchart LR
A[Claim] --> B[Deployment constraint]
B --> C[Measured outcome]
C --> D[Decision]
Local models change privacy and resilience
A robot in a hospital, factory, or home may observe sensitive activity. Sending every frame to a remote service raises privacy, bandwidth, and outage questions. The on-device path can keep more raw data local, but operators still need controls for storage and debugging. The International Federation of Robotics’ service-robot materials show how varied these environments are: https://ifr.org/service-robots. Privacy must be designed around the robot’s actual sensors, not just the model’s marketing category.
A robot in a hospital, factory, or home may observe sensitive activity. Sending every frame to a remote service raises privacy, bandwidth, and outage questions. The on-device path can keep more raw data local, but operators still need controls for storage and debugging. The International Federation of Robotics’ service-robot materials show how varied these environments are: https://ifr.org/service-robots. Privacy must be designed around the robot’s actual sensors, not just the model’s marketing category. The first reporting question is not whether the idea sounds ambitious, but which dependency could make it fail in an ordinary production environment. The practical implication is that readers should separate what the source directly establishes from what follows as analysis. That discipline matters for gemini robotics on-device 2 makes physical ai a data problem: a plausible mechanism is not the same thing as observed performance, and a vendor, research group, or forecaster should be credited for its claim rather than treated as a neutral narrator.
A robot in a hospital, factory, or home may observe sensitive activity. Sending every frame to a remote service raises privacy, bandwidth, and outage questions. The on-device path can keep more raw data local, but operators still need controls for storage and debugging. The International Federation of Robotics’ service-robot materials show how varied these environments are: https://ifr.org/service-robots. Privacy must be designed around the robot’s actual sensors, not just the model’s marketing category. A second question concerns incentives. The organization making the announcement benefits from a favorable interpretation, while customers, workers, or operators carry different risks. Good analysis keeps those positions visible and asks what independent measurement would change the conclusion. The practical implication is that readers should separate what the source directly establishes from what follows as analysis. That discipline matters for gemini robotics on-device 2 makes physical ai a data problem: a plausible mechanism is not the same thing as observed performance, and a vendor, research group, or forecaster should be credited for its claim rather than treated as a neutral narrator.
The interface between model and controller matters
A language-capable policy should not be allowed to directly improvise every motor command. A safer architecture separates semantic planning from constrained motion control, with a supervisor checking workspace, speed, and force limits. ROS 2’s documentation makes the middleware boundary concrete: https://docs.ros.org/en/rolling/. The model can propose a task; deterministic components should decide whether the proposal is physically legal.
A language-capable policy should not be allowed to directly improvise every motor command. A safer architecture separates semantic planning from constrained motion control, with a supervisor checking workspace, speed, and force limits. ROS 2’s documentation makes the middleware boundary concrete: https://docs.ros.org/en/rolling/. The model can propose a task; deterministic components should decide whether the proposal is physically legal. The first reporting question is not whether the idea sounds ambitious, but which dependency could make it fail in an ordinary production environment. The practical implication is that readers should separate what the source directly establishes from what follows as analysis. That discipline matters for gemini robotics on-device 2 makes physical ai a data problem: a plausible mechanism is not the same thing as observed performance, and a vendor, research group, or forecaster should be credited for its claim rather than treated as a neutral narrator.
A language-capable policy should not be allowed to directly improvise every motor command. A safer architecture separates semantic planning from constrained motion control, with a supervisor checking workspace, speed, and force limits. ROS 2’s documentation makes the middleware boundary concrete: https://docs.ros.org/en/rolling/. The model can propose a task; deterministic components should decide whether the proposal is physically legal. A second question concerns incentives. The organization making the announcement benefits from a favorable interpretation, while customers, workers, or operators carry different risks. Good analysis keeps those positions visible and asks what independent measurement would change the conclusion. The practical implication is that readers should separate what the source directly establishes from what follows as analysis. That discipline matters for gemini robotics on-device 2 makes physical ai a data problem: a plausible mechanism is not the same thing as observed performance, and a vendor, research group, or forecaster should be credited for its claim rather than treated as a neutral narrator.
Robot fleets need observability
When a fleet behaves differently across sites, the explanation may be a camera, a firmware version, a payload, or a changed surface rather than the model weights. OpenTelemetry provides a general approach for tracing distributed systems: https://opentelemetry.io/docs/. Robotics teams need an equivalent habit for physical events: record sensor conditions, policy output, controller decision, and human intervention. Without that chain, incident review becomes anecdote.
When a fleet behaves differently across sites, the explanation may be a camera, a firmware version, a payload, or a changed surface rather than the model weights. OpenTelemetry provides a general approach for tracing distributed systems: https://opentelemetry.io/docs/. Robotics teams need an equivalent habit for physical events: record sensor conditions, policy output, controller decision, and human intervention. Without that chain, incident review becomes anecdote. The first reporting question is not whether the idea sounds ambitious, but which dependency could make it fail in an ordinary production environment. The practical implication is that readers should separate what the source directly establishes from what follows as analysis. That discipline matters for gemini robotics on-device 2 makes physical ai a data problem: a plausible mechanism is not the same thing as observed performance, and a vendor, research group, or forecaster should be credited for its claim rather than treated as a neutral narrator.
When a fleet behaves differently across sites, the explanation may be a camera, a firmware version, a payload, or a changed surface rather than the model weights. OpenTelemetry provides a general approach for tracing distributed systems: https://opentelemetry.io/docs/. Robotics teams need an equivalent habit for physical events: record sensor conditions, policy output, controller decision, and human intervention. Without that chain, incident review becomes anecdote. A second question concerns incentives. The organization making the announcement benefits from a favorable interpretation, while customers, workers, or operators carry different risks. Good analysis keeps those positions visible and asks what independent measurement would change the conclusion. The practical implication is that readers should separate what the source directly establishes from what follows as analysis. That discipline matters for gemini robotics on-device 2 makes physical ai a data problem: a plausible mechanism is not the same thing as observed performance, and a vendor, research group, or forecaster should be credited for its claim rather than treated as a neutral narrator.
The economic promise is task density
Robots do not become valuable because they can perform a long list of actions in a lab. They become valuable when they complete enough useful tasks per hour to justify supervision, maintenance, and downtime. The IFR World Robotics reports offer context for the current industrial base: https://ifr.org/worldrobotics. On-device models may improve economics by reducing network dependence, but only if reliability rises with capability.
Robots do not become valuable because they can perform a long list of actions in a lab. They become valuable when they complete enough useful tasks per hour to justify supervision, maintenance, and downtime. The IFR World Robotics reports offer context for the current industrial base: https://ifr.org/worldrobotics. On-device models may improve economics by reducing network dependence, but only if reliability rises with capability. The first reporting question is not whether the idea sounds ambitious, but which dependency could make it fail in an ordinary production environment. The practical implication is that readers should separate what the source directly establishes from what follows as analysis. That discipline matters for gemini robotics on-device 2 makes physical ai a data problem: a plausible mechanism is not the same thing as observed performance, and a vendor, research group, or forecaster should be credited for its claim rather than treated as a neutral narrator.
Robots do not become valuable because they can perform a long list of actions in a lab. They become valuable when they complete enough useful tasks per hour to justify supervision, maintenance, and downtime. The IFR World Robotics reports offer context for the current industrial base: https://ifr.org/worldrobotics. On-device models may improve economics by reducing network dependence, but only if reliability rises with capability. A second question concerns incentives. The organization making the announcement benefits from a favorable interpretation, while customers, workers, or operators carry different risks. Good analysis keeps those positions visible and asks what independent measurement would change the conclusion. The practical implication is that readers should separate what the source directly establishes from what follows as analysis. That discipline matters for gemini robotics on-device 2 makes physical ai a data problem: a plausible mechanism is not the same thing as observed performance, and a vendor, research group, or forecaster should be credited for its claim rather than treated as a neutral narrator.
What “general-purpose” should mean
A general-purpose robot should not be defined by the number of demonstrations in a video. A better definition includes how quickly it adapts, how clearly it signals uncertainty, how safely it recovers, and how much operator training is required. DeepMind’s model cards are useful because they expose intended use and evaluation boundaries rather than treating a model name as a capability guarantee.
A general-purpose robot should not be defined by the number of demonstrations in a video. A better definition includes how quickly it adapts, how clearly it signals uncertainty, how safely it recovers, and how much operator training is required. DeepMind’s model cards are useful because they expose intended use and evaluation boundaries rather than treating a model name as a capability guarantee. The first reporting question is not whether the idea sounds ambitious, but which dependency could make it fail in an ordinary production environment. The practical implication is that readers should separate what the source directly establishes from what follows as analysis. That discipline matters for gemini robotics on-device 2 makes physical ai a data problem: a plausible mechanism is not the same thing as observed performance, and a vendor, research group, or forecaster should be credited for its claim rather than treated as a neutral narrator.
A general-purpose robot should not be defined by the number of demonstrations in a video. A better definition includes how quickly it adapts, how clearly it signals uncertainty, how safely it recovers, and how much operator training is required. DeepMind’s model cards are useful because they expose intended use and evaluation boundaries rather than treating a model name as a capability guarantee. A second question concerns incentives. The organization making the announcement benefits from a favorable interpretation, while customers, workers, or operators carry different risks. Good analysis keeps those positions visible and asks what independent measurement would change the conclusion. The practical implication is that readers should separate what the source directly establishes from what follows as analysis. That discipline matters for gemini robotics on-device 2 makes physical ai a data problem: a plausible mechanism is not the same thing as observed performance, and a vendor, research group, or forecaster should be credited for its claim rather than treated as a neutral narrator.
The next proof is a recovery trace
The most persuasive evidence will be public traces of recovery: what the robot saw, what it attempted, when it stopped, and how a human resumed the task. Researchers should publish distributions of success and intervention, not only best-case videos. If Gemini Robotics On-Device 2 makes that evidence easier to collect locally, its contribution will extend beyond a model release. It will help turn physical AI from spectacle into an engineering discipline.
The most persuasive evidence will be public traces of recovery: what the robot saw, what it attempted, when it stopped, and how a human resumed the task. Researchers should publish distributions of success and intervention, not only best-case videos. If Gemini Robotics On-Device 2 makes that evidence easier to collect locally, its contribution will extend beyond a model release. It will help turn physical AI from spectacle into an engineering discipline. The first reporting question is not whether the idea sounds ambitious, but which dependency could make it fail in an ordinary production environment. The practical implication is that readers should separate what the source directly establishes from what follows as analysis. That discipline matters for gemini robotics on-device 2 makes physical ai a data problem: a plausible mechanism is not the same thing as observed performance, and a vendor, research group, or forecaster should be credited for its claim rather than treated as a neutral narrator.
The most persuasive evidence will be public traces of recovery: what the robot saw, what it attempted, when it stopped, and how a human resumed the task. Researchers should publish distributions of success and intervention, not only best-case videos. If Gemini Robotics On-Device 2 makes that evidence easier to collect locally, its contribution will extend beyond a model release. It will help turn physical AI from spectacle into an engineering discipline. A second question concerns incentives. The organization making the announcement benefits from a favorable interpretation, while customers, workers, or operators carry different risks. Good analysis keeps those positions visible and asks what independent measurement would change the conclusion. The practical implication is that readers should separate what the source directly establishes from what follows as analysis. That discipline matters for gemini robotics on-device 2 makes physical ai a data problem: a plausible mechanism is not the same thing as observed performance, and a vendor, research group, or forecaster should be credited for its claim rather than treated as a neutral narrator.
The reporting standard that matters
The strongest way to follow this story is to watch the evidence trail rather than the announcement cycle. Primary documents establish what was released, promised, or measured. Independent evaluations establish whether the result generalizes. Operational data establishes whether people can use it repeatedly. Those are different forms of evidence, and collapsing them into one headline is how AI coverage becomes less useful than the systems it describes.
Sources consulted
- https://deepmind.google/models/model-cards/gemini-robotics-on-device-2/
- https://deepmind.google/blog/gemini-robotics-er-1-6/
- https://deepmind.google/blog/scaling-up-learning-across-many-different-robot-types/
- https://developer.nvidia.com/isaac
- https://developer.nvidia.com/isaac/sim
- https://docs.ros.org/en/rolling/
- https://opentelemetry.io/docs/
- https://ifr.org/worldrobotics
- https://www.automate.org/robotics/safety
- https://ifr.org/service-robots