Cloud Control of Robots: A 36-Billion-Parameter Model and Latency Under 50 ms

In the demonstration reviewed, Percepton’s IA 05 model with 36 billion parameters was used to control robots in real time. The main computations were performed on remote servers. The demonstration reported latency under 50 ms and operation without a local GPU capable of running a model of that size. This is the result of one experiment, not a specification that applies to every system.

Why Robots Need to Respond Quickly

The description of the demonstration does not reveal what actions the robots performed or how control quality was assessed. It also contains no data that would allow judging what latency is acceptable for other robots or tasks.

A single latency figure does not answer whether a system will suit a particular application. Assessing that requires test results in the target environment and a description of exactly how response time was measured.

How a Large Model Works Through the Cloud

In the demonstration reviewed, IA 05 ran through the cloud, with the main computations performed on remote servers. According to the demonstration, the robot did not need a local GPU capable of running a model with 36 billion parameters.

The available description does not specify what tasks the robot performed, how data was transferred between it and the servers, or what happened when the connection failed. These details cannot be assumed to be established for IA 05.

If cloud control is being considered for another system, this example cannot predict how it will behave when a response is delayed or the connection is lost. The description of the demonstration does not provide such results.

What Latency Under 50 ms Shows

The demonstration reviewed reported that latency was kept below 50 ms. The test conditions and measurement method are not described in the available information.

It is therefore unknown what exactly this figure included: computation time, command transmission, or the full operating cycle. Without these details, the result cannot be compared with tests of another system.

The claim of latency under 50 ms applies only to the experiment described. It does not confirm that this figure would hold under other conditions.

To assess the result for a particular robot, information about the measurement method and tests in the target environment would be needed. The description of the demonstration does not include it.

Risks of Network Dependence

The available information about IA 05 does not describe how the system behaves when a connection is delayed or lost. The consequences of a network failure cannot be assessed from this demonstration.

What happens if the network connection is lost? The experiment description does not answer this. Assessing a particular system requires separate information about its behavior in that situation.

It is also unspecified what data was sent to the server and who could view the logs. The demonstration description does not provide enough information to assess data security and privacy.

The information presented concerns the IA 05 demonstration and does not allow the suitability of cloud control to be assessed for all robots.

First-Person Video and Training Data Collection

The description of the IA 05 demonstration does not say whether first-person video was used to train the model. It also provides no information about how this approach compares with teleoperation recordings.

Therefore, this example cannot establish whether first-person video replaces teleoperated demonstrations or serves as an additional data source.

The available information does not address whether results obtained from video can be transferred to a particular robot. The applicability of this approach to a specific task was not assessed here.

The description of the demonstration contains no data comparing approaches to training IA 05.

When Cloud Control Makes Sense

A single result is not enough to determine whether a cloud model is suitable for a particular robot. If such a system is being considered, it is important to compare its test conditions with the target environment and find out how it responds to delays or a lost connection. The description of IA 05 does not provide this information.

• Find out how the reported latency of under 50 ms was measured.
• Compare the demonstration conditions with the robot’s operating environment, if those conditions are known.
• Find out how the system responds to delays or a lost connection. The description of IA 05 does not reveal this.
• Test the model on the actions the robot needs to perform for the target task.
• Find out what data is sent to the server: the available information about IA 05 does not specify this.

In the demonstration reviewed, the IA 05 model with 36 billion parameters was used to control robots via the cloud, with latency reported at under 50 ms. The description does not reveal the measurement method, test conditions, or behavior during connection failures, so this result cannot be used to assess the system’s performance in a different environment.

Frequently Asked Questions About Cloud Control of Robots

Can a robot be controlled through a cloud model?

In the demonstration reviewed, IA 05 was used to control robots in real time. This example confirms only that a demonstration took place, not that the model is suitable for other tasks.

Does a robot need a powerful local GPU?

In the demonstration, the main computations were performed on remote servers, and a local GPU capable of running a model of that size was not required. Other local system components are not described in the available information.

Does first-person video replace teleoperated demonstrations?

The description of the IA 05 demonstration does not confirm this. It does not say whether first-person video was used to train the model.