Target-driven Visual Navigation in Indoor Scenes using Deep Reinforcement Learning



in this video we introduce our work target driven visual navigation in indoor scenes using deep reinforcement learning we developed a deep reinforcement learning algorithm based only on visual input to navigate in a space to find a given target training deep reinforcement learning methods in the real environment is often infeasible due to their demand for time and data hence we train our algorithm in a simulated world and later adapted to the real world we developed one of the first simulation frameworks with high quality 3d themes called AI to Thor our framework enables a rich set of physical interactions with objects the agent can navigate in the environment apply forces to objects and change their state standard deep reinforcement learning methods take only the observation is input therefore they need to be retrained for new goals in contrast our deep reinforcement learning model learns a function of both the current observation and the target hence it eliminates the need for retraining for new targets which results in generalization across targets and scenes this is a training episode for the task of finding a sofa the model initially performs a random walk in the top right corner you see a bird's-eye view of the scene the model gradually converges as it observes more frames during training here is the final result of our model here is another example that shows navigation toward books our deep reinforcement learning model learns to navigate to new targets that are not used during training we also show that the model can navigate to find targets in a new scene we also show the extension of our model to continuous base this involves physical interaction and handling noisy movements we also show that the model that is trained in a simulated environment can be used in a real setting after minimal fine-tuning we move the robot in the scene and collect images for fine-tuning the model we specify the door as the target and the robot learns to navigate to the door on the right side you can see a view of the robot here is another example that shows navigation toward the microwave for quantitative evaluations please refer to our paper

1 thought on “Target-driven Visual Navigation in Indoor Scenes using Deep Reinforcement Learning”

Leave a Reply

Your email address will not be published. Required fields are marked *