Optimizing Learning for Robotics

We are building a system based on ROS that non-invasively enables coordinated and collision-free navigation of multiple autonomous robots. Our system makes no assumption about the inner workings of the robot's path-planner or motion controller. We use our system to enable a heterogeneous fleet of autonomous robots to safely navigate a shared environment.

We are building a ROS local planner that's driven by a Reinforcement Learning (RL) agent. the agent state is an image derived from the local dynamic environment and the training is similar to RL training for agents. The resulting agent controls the robot from the cloud and enables safe navigation in dynamic environments with unpredictable obstacles (such as people working in a factory or warehouse).

We localize indoor robots using camera-tracing and noisy RF sensors. Our algorithms employ robot velocity prediction and ongoing correlation against original robot plans.