When will humanoid robots be safe enough for the manufacturing floor?

Humanoids are in a development mode, and many wonder when the reality of humanoids will affect manufacturing. Is it reality or science fiction? Boston Dynamics’ humanoid is Atlas. Atlas has been under development since 2013. In 2014, it was a Defense Advanced Research Projects Agency (DARPA) project. One of the drawbacks has been starting with hydraulic actuators.

In 2024, using 3D components, the Atlas system moved to all electric actuators. Some say that Atlas is the most advanced humanoid in existence, but the discussion here is not about that. I listened to the Boston Dynamics webinar, “Humanoids and the Unbound Factory.”

Hardware, software, dexterity

The discussion was about whether the humanoid software is keeping up and that the hardware is advanced enough that humanoids can do the tasks required for manufacturing repeatable tasks and do so safely and reliably.

What does this involve? Humanoids must have the dexterity to do the type of work that humans do. Boston Dynamics breaks this down into hardware, software, behavior, safety and reliability.

Material science has advanced to allow grasping and whole-body tactile sensing. Reinforcement and imitation learning are a way of teaching dexterity. Advances in hardware allow for more dexterity.

To compile the inputs of all the sensors to allow for a humanoid to process its environment, software is key. Artificial intelligence should allow us to manage this. The hardware and software are constantly changing, but hardware is to a point that human-like movement has been obtained.

Behavioral modeling

Behavior and safety are more complex than the technology, in my opinion. Why? How do you model behavior that allows a robot to do a task? I say model, because tasks should be adaptable and the goals in the humanoid arena are human interfaces that allow the robot to be taught by vision and showing. These are the same goals used to teach humans a repetitive process.

However, ask a mom how to teach a child to ride a bicycle, and you will get more than one answer. The key is simplicity and repeatability in a mathematical form that can be turned to 1s and 0s and processed. This is not easy when engineers think it should happen in milliseconds or microseconds. Then, add environment, and process.

How do you scale to that? How do you maintain it? Can the humanoid repeat the task like a human can?

Training methodologies

Boston Dynamics calls this post-training. Post-training is where the robot is shown how to do a task, and then it repeats the task to perfection. Scalability becomes a problem, if this task is tedious and long.

The result was to add pre-training, where the robot absorbs environment sensors and feedback to be able to deploy a feed-forward type of decision-making process, so that it can develop some intuitiveness based on pre-training smaller movements.

This means teaching the robot tasks that are smaller components of large multiple tasks that are a conglomerate of a set of tasks—kind of like crawling before walking or learning how to grip a baseball before throwing it. This relates to visual spatial understanding.

Picture having dexterity, understanding force based on task and then utilizing input from visual spatial understanding and applying that to an environment with tools, machines or people, so that the robot may understand a safety envelope.

Boston Dynamics is using learning from a demonstrator to teach Atlas movements. Teleoperators use technology to see the world through the humanoid eyes and technology on their wrists and ankles to understand how the robot is in space. Then they teach movements.

This is a type of pre-training that is allowing the robot to learn micromovements that will allow full movements in tasks, like a digital twin of the human.

Safety Protective and Assistive Robot Kit

Dexterity, movement and environment, as well as how a robot behaves in an environment, are directly related to safety. New frameworks are under development in humanoids that are applicable to any robot working in manufacturing.

Get your subscription to Control Design’s daily newsletter.

Safety Protective and Assistive Robot Kit (SPARK) is a new modular safety control framework. “SPARK: A Modular Benchmark for Humanoid Robot Safety,” a paper written by Yifan Sun, Rui Chen, Kai S. Yun, Yikuan Fang, Sebin Jung, Feihan Li, Bowei Li, Weiye Zhao, Changliu Liu, is available on arXiv.

SPARK uses middleware that is compatible with a robot operating system (ROS). Also, there is a data distribution system (DDS) that is usable for the robot safety system to process environment and movements. The idea behind a safety framework is to process the application with the humanoid systems.

NexCom has been working on system hardware and software interfaces that allow the safety framework to be run on a separate controller from the robot operating system but to provide a constant interface. This is important for behavior because the amount of data to move in an environment is just as much as the data being processed to move.

Functional safety on the factory floor

This is the same idea as what the machine builder is doing with the safety programmable logic controller (PLC). How so? A separate process is looking at the environment and can put robot or system behavior in check if it sees an unsafe action in the monitored environment.

Why is it important to put systems, hardware and sensor inputs into a framework that can be processed separately from process, while interacting with process? Humanoids are still restricted, because of functional safety. There is no certification in place for functional safety of a humanoid robot.

What is the difference between functional safety for a moving system and a static machine that only has state changes around it? The humanoid is moving through its environment and having things move around it, while doing a task. The static machine only has the one complexity of task interacting with the humans moving around it.

But what if your machine is as long as a football field? See how putting a separate safety processor in play works to your advantage? Thus, we get the exchange from research to the factory floor.

Manufacturing implications

How does advancing humanoids help us on the manufacturing floor? It makes our automation more developed. Figuring everything out in the research lab helps us get better hardware on the plant floor. This means, manufacturing advances through better technology, applied to machines that are built around applications, not technology.

Humanoid builders are still at Level 1 in design phase, but they want to be able to sell the humanoid as a repeatable, safe, reliable, manufacturing tool that can enable customers to progress with their own goals.

Maybe, in the future, original equipment manufacturers (OEMs) can build a machine that comes with its own operator. Why else would you build a humanoid? Perhaps if a robot were running a machine, then the arguments about why the machine did better in manual mode than automatic would go away.