{"about":{"site":"https://codewithpapers.app","non_affiliation":"Code with Papers and Syntology are not affiliated with, endorsed by, or sponsored by Papers with Code, Meta, or the pwc-archive mirror.","licence":"CC BY-SA 4.0","licence_url":"https://creativecommons.org/licenses/by-sa/4.0/legalcode","attribution":"https://codewithpapers.app/attribution","modified":"archive material modified by Syntology; see the attribution page"},"url":"/paper/cad2rl-real-single-image-flight-without-a","title":"CAD2RL: Real Single-Image Flight without a Single Real Image","arxiv_id":"1611.04201","date":"2016-11-13","proceeding":null,"authors":["Fereshteh Sadeghi","Sergey Levine"],"abstract":"Deep reinforcement learning has emerged as a promising and powerful technique\nfor automatically acquiring control policies that can process raw sensory\ninputs, such as images, and perform complex behaviors. However, extending deep\nRL to real-world robotic tasks has proven challenging, particularly in\nsafety-critical domains such as autonomous flight, where a trial-and-error\nlearning process is often impractical. In this paper, we explore the following\nquestion: can we train vision-based navigation policies entirely in simulation,\nand then transfer them into the real world to achieve real-world flight without\na single real training image? We propose a learning method that we call\nCAD$^2$RL, which can be used to perform collision-free indoor flight in the\nreal world while being trained entirely on 3D CAD models. Our method uses\nsingle RGB images from a monocular camera, without needing to explicitly\nreconstruct the 3D geometry of the environment or perform explicit motion\nplanning. Our learned collision avoidance policy is represented by a deep\nconvolutional neural network that directly processes raw monocular images and\noutputs velocity commands. This policy is trained entirely on simulated images,\nwith a Monte Carlo policy evaluation algorithm that directly optimizes the\nnetwork's ability to produce collision-free flight. By highly randomizing the\nrendering settings for our simulated training set, we show that we can train a\npolicy that generalizes to the real world, without requiring the simulator to\nbe particularly realistic or high-fidelity. We evaluate our method by flying a\nreal quadrotor through indoor environments, and further evaluate the design\nchoices in our simulator through a series of ablation studies on depth\nprediction. For supplementary video see: https://youtu.be/nXBWmzFrj5s","url_abs":"http://arxiv.org/abs/1611.04201v4","url_pdf":"http://arxiv.org/pdf/1611.04201v4.pdf","source":{"archive":"pwc-archive (Hugging Face), CC BY-SA 4.0","snapshot":"2025-07-28","licence_url":"https://creativecommons.org/licenses/by-sa/4.0/legalcode","row_kind":"abstracts"},"code_links":[{"paper_slug":"cad2rl-real-single-image-flight-without-a","repo_url":"https://github.com/abefetterman/hamstir-gym","is_official":0,"mentioned_in_paper":0,"mentioned_in_github":1,"framework":"tf","reach":{"status":"ok","spdx":"MIT"}}],"tasks":[{"task_slug":"3d-geometry","task_name":"3D geometry"},{"task_slug":"collision-avoidance","task_name":"Collision Avoidance"},{"task_slug":"deep-reinforcement-learning","task_name":"Deep Reinforcement Learning"},{"task_slug":"depth-estimation","task_name":"Depth Estimation"},{"task_slug":"depth-prediction","task_name":"Depth Prediction"},{"task_slug":"motion-planning","task_name":"Motion Planning"},{"task_slug":"reinforcement-learning","task_name":"Reinforcement Learning"}],"methods":[],"datasets_introduced":[],"methods_introduced":[],"results":[],"syntology":{"atlas_url":"https://app.syntology.ai/?focus=1611.04201","mcp":null,"developers":"https://syntology.ai/developers"},"arxiv_metadata":null,"syntology_extracted_results":null}