Enabling Reinforcement Learning on Industrial Robots Using ForgeOS
The field of industrial robotics is on the cusp of a significant transformation, as machine learning and Artificial Intelligence-based methods are paving the way for robots to accomplish more complex tasks. Reinforcement learning (RL) is a specific area that promises more advanced capabilities, especially in how robots can achieve goals through their motion.
Two issues are blocking the wide adoption of RL on industrial robots for manufacturing tasks. The first bottleneck is the absence of standardized, low-latency streaming joint interfaces on industrial robots. This type of joint interface is particularly crucial for implementing motion policies derived from reinforcement learning (RL) because of the tight feedback required between the command to the robot and information about its behavior. Unfortunately, few robot brands support this interface, the interfaces for those that do are all different, and no industrial-hardened software has provided a consistent interface until now.
The second issue has been a lack of high-level task orchestration tools that can work alongside RL-learned behaviors. On most RL applications we have seen on industrial robots the learned behavior, while impressive, was run in isolation of a full user-specified program, without coordination with other tools or devices needed for industrial tasks. READY’s ForgeOS and its Robot Abstraction Layer have been a breakthrough for agnostic robot control. We feel that similar agnosticism around joint streaming for robots, combined with Forge’s exceptional orchestration and task management capabilities will open the applications of RL to a much larger install base.