Dr. Florian Golemo

Real guide dogs are incredibly expensive (50,000 USD) and take long to train. With Milo, we’ve developed an autonomous (self-contained, no cloud compute) and low-cost (2,000 USD) robot guide dog that can follow paths and avoid obstacles and pedestrians indoors and outdoors. And in order to foster open access for the blind/low-vision community, we’re open-sourcing our research.

Domain randomization is a popular technique for improving domain transfer, often used in a zero-shot setting when the target domain is unknown or cannot easily be used for training. In this work, we empirically examine the effects of domain randomization on agent generalization. Our experiments show that domain randomization may lead to suboptimal, high-variance policies, which we attribute to the uniform sampling of environment parameters. We propose Active Domain Randomization, a novel algorithm that learns a parameter sampling strategy. Our method looks for the most informative environment variations within the given randomization ranges by leveraging the discrepancies of policy rollouts in randomized and reference environment instances. We find that training more frequently on these instances leads to better overall agent generalization. Our experiments across various physics-based simulated and real-robot tasks show that this enhancement leads to more robust, consistent policies.

Despite the recent successes of deep reinforcement learning, teaching complex motor skills to a physical robot remains a hard problem. While learning directly on a real system is usually impractical, doing so in simulation has proven to be fast and safe. Nevertheless, because of the "reality gap," policies trained in simulation often perform poorly when deployed on a real system. In this work, we introduce a method for training a recurrent neural network on the differences between simulated and real robot trajectories and then using this model to augment the simulator. This Neural-Augmented Simulation (NAS) can be used to learn control policies that transfer significantly better to real environments than policies learned on existing simulators. We demonstrate the potential of our approach through a set of experiments on the Mujoco simulator with added backlash and the Poppy Ergo Jr robot. NAS allows us to learn policies that are competitive with ones that would have been learned directly on the real robot.