Synthiam Exosphere combines AI, human operators to train robots
Most robots are as good as or even better than humans at certain repetitive tasks. However, autonomous systems operating in complex, dynamic environments can benefit from human judgment in edge-case situations. Synthiam Inc. has developed Exosphere, which bids out tasks for cloud-based robots to artificial intelligence and human controllers.
Calgary, Alberta-based Synthiam also offers Autonomous Robot Control (ARC) software and a Skill Store for robotics programming. The startup claims to have helped develop more than 5,800 robot apps, 2.6 million connections, and 472 behavior controls.
Dealing with data and coding
“My previous company was based on robot modularity and is in the niche market of education,” said DJ Sures, who founded EZ-Robot Inc. in 2012. He sold it and founded Synthiam in 2019. “When Synthiam released the Microsoft Cognitive Services robot skill, they thought they got hacked because so much data was coming in.”
“ROS is great for programmers and hobbyists,” asserted Sures, referring to the Robot Operating System. “What we’re doing is for 99% of the rest of the world that doesn’t want to get into programming. We had the idea of hands-off execution.”
Users of Synthiam’s products include Microsoft and NASA, he told The Robot Report.
“There are tons of ‘if’ conditions now in laboratories,” said Sures. “With a linear model or ideology, the programmer has to compensate for every problem that emerges. But merging those lessons with AI is not quite working. Exosphere can now bridge the valley between autonomy and needing a user. It takes environmental data and data from simulation and adds human intelligence for machine learning.”
Exosphere crowdsources intelligence in the cloud
“Exosphere was designed to do three things,” Sures explained. “First, use human operators to perform the more than 10% of edge cases that robots can’t do. Second is to train the AI and machine learning [ML] knowledge base of visual and decision data so robots can get smarter. Third is to provide access to real robots for AI/ML companies to test algorithms in the real world versus the lab.”
“Exosphere is a transactional service above the cloud, connecting AI, robots, and humans,” he added. “It sends off a contract, rated by degree of confidence, and then to human controllers for crowdsourced intelligence. People control the robot, contributing to the data set, which is re-run in simulation to improve the algorithms.”
“This also helps people understand how robots function in the real world,” said Sures. “Just as Unity enabled creators to easily build video games, we want to use humans to help build the data sets to train robots.”
“We’re working hard on security,” he added. “Operators are anonymized — we use facial tracking to blur faces and to allow for safety. The person who asks the robot for a task has a panic button. After the task has been done, there’s a reporting button, and both sides validate performance and rate the task.”
Working with robots
Synthiam is working with robotics and AI companies to build training data. “We’ve had no challenge finding robots to use with our systems,” Sures said. “Most are humanoid; some are flying. The software doesn’t care what type of locomotion a robot uses. The software is like a translator for how the robot moves.”
“One robotics startup was spending 75% of its efforts on anything but AI, for which it must create an environmental simulation,” said Sures. “In other cases, AI departments spend a lot of effort getting robots to talk to them. We’re looking at giving a critical mass of small robotics companies access to AI.”
“This would allow developers to access robots in homes and random environments,” he added. “All have robot definition files, allowing human intelligence, or HI, to train and oversee AI. Investors can test if an AI works in Exosphere.”
Use cases for Exosphere
Many providers of autonomous mobile robots (AMRs) for warehouses and retail have central operations centers to handle frequent edge cases, Sures said.
“Whenever AMRs have to interact with a person, pilots can fail,” he said. “A level of abstraction is missing. What is the customer’s problem? Our software is the compatibility layer — it doesn’t care how you move. Think of it like a concierge service in a hotel.”
“It doesn’t matter to the customer if it’s AI or a human that’s controlling a robot, and you can have one person controlling 10 robots,” said Sures. “We recently demonstrated a ‘playground,’ where four robots were controlled for certain basic tasks. The robots for the playground were actual robots that people used to train with before controlling real robots.”
The COVID-19 pandemic has increased interest in developing remotely controlled and more autonomous robots for delivering food and medicine, disinfection, and other uses. However, many are still in trials.
“The reason we’re not seeing a lot of robots assisting with COVID-19 is because of the edge cases that make them unreliable in critical situations,” Sures said. “Exosphere-powered robots can successfully complete any task by leveraging human operators for as little or as much as is necessary. In turn, they continue training the knowledge base to make the robots smarter for next time.”
Synthiam Exosphere is compatible with a wide range of platforms and is now live in community beta.