Presented by: Hones Approach to Predictive Maintenance, Says RoboBusiness Panelist


September 18, 2018      

Just as robotics has made production more efficient, so too can artificial intelligence help reduce downtime through predictive maintenance. Startups aiming to serve this market must first understand the capabilities of their technology and where the value proposition lies.

Bend, Ore.-based recently refocused on better serving manufacturers. Stephen Sickler, who became president of this February, explained this shift to Robotics Business Review.

Sickler was a founder of the MSPAlliance, the largest and oldest vendor-neutral organization for cloud and managed service providers. He has also held senior management positions at leading software-as-a-service companies.

Here, Sickler gives a preview of his session at RoboBusiness and describes how Tend learned to better serve its target market via predictive maintenance.

Q: What brings you to robotics and artificial intelligence?

Sickler: In my career, I’ve worked at a series of startups. I was at Tivoli, which was bought by IBM. Then I was at Vantive, and Oracle bought PeopleSoft. Then I worked at the MSPAlliance and SiteLite.'s Stephen Sickler

Stephen Sickler will discuss machine learning and predictive maintenance at RoboBusiness 2018.

I live in Bend and was involved in a mentoring organization for startups called Opportunity Knocks or “OK.” More than 30 executives from SAP, IBM, and Oracle live in Bend or nearby.

We recruited people with operational experiences. James Gentes, co-founder of Tend, was in my group, so I heard about Tend before it was called “Tend.”

I mentored James through first couple of phases of Tend’s existence. There has been quite a bit of change [among executives], but the company has now found its market.

Q: So what is all about?

Sickler: The company first started with the bold idea for a unique platform to program lots of different robots from a laptop or a phone.

It went to market — programming robots is a precise exercise, but doing it through the cloud with lag made it more frustrating. That’s a technical problem that the organization couldn’t quite get over.

Systems integrators are really the ones who set up and program robots. Changing their programming is actually pretty rare. The initial premise didn’t have much practical business value.

However, in developing the platform, Tend learned how to talk to a number of different robots and how to generate alerts, texts, or e-mails. This enabled people to connect and analyze the cause of a problem. They could see the state of a robot and use cameras to see the entire workcell.

GE Appliance was Tend’s flagship customer last October, when it launched the product as In.view.

Q: What market are you now focusing on?

Sickler: Come the end of last year, there was a mutual decision with the investors that the company hadn’t gotten the revenue expected. Venture capitalists had spent a lot of money, and Tend had developed a lot of software.

I joined as president and chief operating officer, and I talked with a lot of systems integrators. We realigned to industrial robots.

Collaborative robots are the new shiny object, but we’re really focused on the huge market of industrial robots, which has a large installed base.

Integrators told us two things: Getting information about a robot is helpful and useful, but it’s nice to get a more holistic view of a cell, such as the PLCs [programmable logic controllers] controlling conveyor belts or parts bins.

The second thing — when I visited Toyota in Kentucky, I saw that it produces one Camry every 57 seconds. If a robot has an unforeseen failure, it costs $3 million for yanking a robot arm and putting in a new one.

With the predictive maintenance aspect [of’s platform], we were on to something.

We’ve solved both problems. We now have the capability [with In.production] to get information from PLCs to decipher what’s going on in the cell with robot arms or processes.

Perhaps the robot has stopped, but it could be because of a signal for a new part, or it might be a sensor not making contact.

The second problem will launch us into the future. Tend’s machine-learning algorithm is an anomaly-detection engine. With [In.advance], you can set a baseline of what’s normal based on any data. You can monitor current by servo, temperature, torque, or vibration. It can then detect when something’s not normal and send alerts by variance.

As Chris Daniels, president of integrator Camtech Manufacturing Solutions and a fellow speaker on my panel, can attest, we’ve had some instances where there were no visible signs of a problem. There were no strange noises, but we sensed that current usage was going up, and we found a bearing in a wrist that’s starting to fall apart. This can help avoid catastrophic failure.

Q: What was the first realization of your product’s value for predictive maintenance?

Sickler: Some vendors such as FANUC ZDT [Zero Down Time] ask customers to allow all robot data — an immense amount of data — to be pumped to a data center in Ohio in an attempt at predictive maintenance. Some manufacturers question whether that approach works.'s predictive maintenance product

Tend’s Predictive Maintenance Module tracks torque and can predict equipment failure. Source:

They’re just focused on robot data, like our initial version of In.view, but automakers would like to know how other components, like welding guns or paint sprayers, can be monitored and generate data. That exploded the value equation.

Also, FANUC and ABB have been doing robots for a long time, but their software and tools for customers are primarily bad Windows-based software. That’s understandable at large, older companies that focus on making hardware reliable.

Q: How do you see the robotics market evolving?

Sickler: The U.S. is No. 5 or 6 in robot density. We hear politicians complaining about a lack of competitiveness on labor costs or a lack of environmental rules. I was at a company in Portland, Ore., that makes chainsaws and it now has eight robots.

It’s nice to be able to call GE and get union workers to come fix a part.

Our solution solves two problems: It helps avoid unforeseen robot failure and can give companies the tools to fix day-to-day small issues that affect their workcells. We’re not training robotics engineers fast enough.

We’ve signed three new integrator partners in Detroit. Our new model is now more effective. They’re using it to monitor cells in a 30-to-60-day shakedown period, ramp up the product, and test it.

Q: How closely are you working with integrators?

Sickler: The whole idea is to give integrators tools to improve customer service — that’s part of our value proposition.

We’re about to release an API [application programming interface] that will allow systems integrators to feed us any data stream in a certain format that our anomaly-detection engine can read.

Customers have to send us the data in our format, but that’s all that integrators need to do. People are excited.

Integrators will say there’s an art to programming an industrial robot. One mistake that less-experienced programmers will make is starting and stopping a robot too fast. This creates a lot of wear.

Asking a robot to reach around can put more stress on joints. As soon as we hook up our system, if it sees high torque readings, it can send an alert.

Systems integrators can analyze robot operations for predictive maintenance and tell a customer if a program or a process should be modified to reduce stress.

Camtech’s Daniels has 12 customers that are at various stages of running systems in pilot or production. I’m looking forward to what he’ll have to say at RoboBusiness about results.

Q: What role has Raspberry Pi played at Tend?

Sickler: We’re running an anomaly-detection algorithm on a Raspberry Pi platform. We’re not asking people to share their data like FANUC or Oracle.

We got really lucky that the latest version of Raspberry Pi is suitable for industrial use. It’s UL-, CE-certified. In February, we realized that we really want to do predictive maintenance for Tier 1 automotive suppliers or aerospace manufacturers, so let’s figure out how to do it on-site.

We started testing on a Raspberry Pi platform with four cores, a1.6-GHz processor, and a Micro SD card for memory at pennies. It was a perfect storm.

Raspberry Pi is becoming more powerful. You see this with Google and Amazon research. You can download open-source algorithms basically for free and quickly get to market. We analyzed machine learning on Raspberry Pi. Using our platform to receive alerts passed security audits at GE and Toyota.

There’s so much data coming off these robots, but most are doing repetitive process. Unlike a Universal Robots cobot that’s doing varied movements for pick and pack, industrial robots do the same thing over and over. We’re only going to track current torque and temperature. If you don’t try to analyze or get other data, that greatly simplifies the problem.

Cell.Mate from

Cell.Mate automatically connects to a robot and starts setting a baseline. Source:

Q: Can you explain the difference between “supervised” and “unsupervised” learning?

Sickler: One of our design criteria is we can pop our Cell.Mate into the robot cell. Cell.Mate is on a DIN rail in the PLC cabinet. You plug it into the network with a robot, and it tells us the IP address of the robot. It immediately connects, starts setting a baseline. You run it for three to four days, and there’s no programming to do.

We have standard measurements and alerts for other sensors. We also have a data-entry screen. You can name the sensor and tell us any parameters to measure.

Q: What’s the challenge for the Industrial Internet of Things?

Sickler: Rather than IIoT, it’s lots of data. We had an epiphany with machine learning and predictive maintenance. Data is data — you can get it from cameras or other sensors in a cell.

For example, someone came to us and said their robots don’t fail that often, but the paint gun has a lot of parts that fail much more often. “Can you track that data on pressure, from multiple sensors?”

We’re like, “Sure — baseline normal or not normal.”

We’re super-excited about our platform. It started with robots, but our focus has really turned to anything in a cell that we can collect data on that’s potentially a point of failure. Wireless vibration sensors are a key indicator of failure for a lot of things.

We’ve simplified our approach. Some vendors make AI sound like all data needs to come to a giant data lake, with hundreds of data scientists graduated from MIT to find insights.

Right now, people are saying that — with just our alert data about a stopped robot or something minor like washing-machine tub slides — they can look at where they are losing time cumulatively over a certain part of their process. By doing analysis, they can get out there so it doesn’t do it over and over again.

Over time, we’ll be able to address predictive maintenance with cobots. Everybody wants to get cycle time as fast as possible, with the robot moving at lightning speed, but then it has to wait for next part. That’s not really affecting production speed.

Q: What are your plans for

Sickler: Our long-term vision is to be a unified platform. We have Denso and ABB robots in house as we rushed to become robot-agnostic. In the next few months, we’ll announce additional robots based on market share and customer demand.

We’re in pilots with a number of automotive manufacturers and Tier 1 supporters.

Q: What are you looking forward to at RoboBusiness 2018?

Sickler: I’m looking forward to learning what other companies are doing in this field, generically in AI and robotics. There are possible partnerships to be made.

We’re running into other devices that are in the robot cell that other people are coming up with. Rather than worry about the data format, it’s better to work together.

We’ll show a video of two robots running side by side and how our system can find a problem with one.