MIT Researchers Develop General Optimiser Tool For Any Autonomous Robotic System

Recently, researchers from MIT announced the development of an optimiser tool that can be applied to any autonomous robotic system. What challenges does robotics introduce to engineers, what did the MIT researchers do, and how will the development of generic platforms help to accelerate robotic development?

What challenges does robotics introduce to engineers?

Robotics has undoubtedly transformed the manufacturing industry thanks to the ability of robotics to perform repetitive tasks efficiently and the lack of salary needed by robotic platforms. Furthermore, robotic platforms do not make mistakes through exhaustion (instead, they make mistakes through sensory faults and program bugs), can operate 24/7, and do not require a controlled, comfortable environment with air conditioning. 

While robotics has proven its worth ten-fold in the manufacturing industry, they are yet to enter other industries as a practical platform. Sure, there are robotic systems such as those developed by Boston Dynamics that can perform backflips, or whatever it was that Tesla introduced in their recent AI show. Still, the limitations currently faced by robotics prevent them from being used outside of carefully controlled industrial environments.

However, one of the most prominent challenges engineers face is that robotic platforms are almost always designed from scratch, which not only slows down development but can require unique coding solutions. The lack of generic hardware and software platforms on the market makes it hard for engineers to turn out new robotic designs at speed, and the lack of such platforms also makes it hard for technical standards to form (such as communication buses, hardware fittings, and programming methods). Thus, many engineers will find each new robotics project as a totally new design from the ground up.

Finally, engineers also face enormous challenges when trying to coordinate two different robotic platforms to perform a shared task. As both platforms will undoubtedly have different programming environments using different software interfaces, no one program that describes what each robot needs to do can be written and deployed.

MIT researchers create a robotic optimiser that can target any platform

Recognising the challenges faced by robotic systems, MIT researchers have recently developed a general-purpose optimiser that can help engineers accelerate the time to develop autonomous robotic systems while simultaneously helping engineers coordinate different robotic systems regardless of their form or function. In their demonstration, MIT noted the challenges faced by engineers when designing a robotic system for a specific purpose, including the need for countless trial-and-error simulations, which result in unique solutions for each robotic system developed.

The solution offered by MIT researchers approaches the automation problem by working backwards instead of forwards. Instead of starting with a design and running multiple simulations to find the best solution, the researchers instead started with the ideal solution and then worked backwards to find the optimal design and function. This approach makes sense when considering that it is likely for a roboticist to know what an ideal solution would be but doesn’t know how to make their system behave in that way.

In order to realise the general-purpose optimiser, the researchers utilised a technical called automatic differentiation which has close ties to neural network training. The ability of the optimiser to make small tweaks to the parameters of each robotic system allows for further performance improvement. In their demonstration, the researchers demonstrated two different scenarios, the first involving two robots moving a heavy box into a specific position and the second involving a robot avoiding a box and finding a new path.

How will generic platforms help robotic developers?

What the MIT researchers demonstrated with their optimisation tool is the ability for a complex operation that is well understood by humans to be converted into actions that robotic systems understand. Instead of trying to identify every possible outcome depending on how each robot is moved, the optimiser can instead be told what we want the robots to do so that it can find the best movement parameters for those robots.

However, the development of generic platforms and optimisers also introduces numerous possibilities to engineers when both designing robotic systems and their eventual implementation. Firstly, generic platforms allow engineers to rapidly design new robotic systems while recycling older designs (i.e., stop reinventing the wheel when designing robots). 

Secondly, the use of common platforms also allows engineers to reuse code and more easily integrate different robotic solutions together. This is especially helped with the introduction of platforms such as the Robot Operating System, which provides software libraries and tools to create robotic systems.

Thirdly, the use of unified designs and interfaces can also lead to robotics with interchangeable parts that do not require massive amounts of recoding. For example, an interface bus (such as USB) could allow robotic systems to hot-swap limbs and tools, enabling dynamic operation and allowing robotic systems to be constructed using parts from numerous different manufacturers.

Overall, the robotics industry faces a major challenge with the lack of general-purpose tools and hardware options. Sure, there are many free designs available to engineers, and most robotic forms have been solved but trying to program different robotic systems that can work together is no small feat.