OpenAI published research on 11 October 2016 describing a method to transfer robotic control policies from simulation to physical hardware by learning an inverse dynamics model [source]. The work addressed a persistent challenge in robotics: policies trained in simulation often fail when deployed on real robots due to discrepancies in physics modelling, sensor noise, and actuator dynamics.

The researchers trained a neural network to predict motor commands from desired state changes, effectively learning the inverse of the robot's dynamics. By collecting data from the physical robot and training this model, they created a mapping that compensated for simulation-to-reality gaps. Policies trained entirely in simulation could then be executed on real hardware by passing actions through the learned inverse model.

The team demonstrated the approach on a manipulation task using a robotic arm. A policy trained in MuJoCo simulation was transferred to a physical Fetch robot without additional real-world training of the policy itself. The inverse dynamics model required approximately 1.5 hours of real-world data collection to train.

This work represented an early effort by OpenAI to bridge the simulation-reality divide in robotics, a problem that has remained central to deploying learned control systems. The approach reduced but did not eliminate the need for real-world data, as the inverse model itself required physical interaction to learn. The method assumed that simulation could provide sufficient task structure, with the inverse model correcting for execution-level differences.

The research predated OpenAI's later focus on large language models and marked a period when the organization actively published robotics work. No commercial product resulted directly from this research.