Stanford AA289/ENGR319 - Robotics and Autonomous Systems Seminar - Stanford Seminar - Distributed Perception and Learning Between Robots and the Cloud

Sandeep Chinchali Stanford University January 10, 2020 Today’s robotic fleets are increasingly facing two coupled challenges. First, they are measuring growing volumes of high-bitrate video and LIDAR sensory streams, which, second, requires them to use increasingly compute-intensive models, such as deep neural networks (DNNs), for downstream perception or control. To cope with such challenges, compute and storage-limited robots, such as low-power drones, can offload data to central servers (or “the cloud”), for more accurate real-time perception as well as offline model learning. However, cloud processing of robotic sensory streams introduces acute systems bottlenecks ranging from network delay for real-time inference, to cloud storage, human annotation, and cloud-computing cost for offline model learning. In this talk, I will present learning-based approaches for robots to improve model performance with cloud offloading, but with minimal systems cost. For real-time inference, I will present a deep reinforcement learning based offloader that decides when a robot should exploit low-latency, on-board computation, or, when highly uncertain, query a more accurate cloud model. Then, for continual learning, I will present an intelligent, on-robot sampler that mines real-time sensory streams for valuable training examples to send to the cloud for model re-training. Using insights from months of field data and experiments on state-of-the-art embedded deep learning hardware, I will show how simple learning algorithms allow robots to significantly transcend their on-board sensing and control performance, but with limited communication cost. View the full playlist: https://www.youtube.com/playlist?list=PLoROMvodv4rMeercb-kvGLUrOq4HR6BZD 0:00 Introduction 0:37 Robot sensory data + compute models are becoming increasingly complex 2:03 How Can Network Connectivity Help Robots? 4:43 Key Challenges of Cloud Robotics 6:01 1. Distributed Inference: The Robot-Cloud Offloading Problem 7:26 2. Distributed Learning: The Robot Sensory Sampling Problem 9:31 Outline 10:47 Accuracy of Robot and Cloud DNNS 13:41 Hidden Costs of Network Congestion 13:44 Network Costs of Cloud Communication 14:58 Our Network Congestion Experiments 17:34 Cloud Offloading: A Dynamic Decision-Making Problem 18:10 Robot-Cloud Offloading: Sequential Model Selection 21:15 Reinforcement Learning (RL) 24:02 The Robot Offloading MDP: Action Space 24:25 The Robot Offloading MDP: State Space 25:55 The Robot Offloading MDP: Reward 27:28 Deep RL beats benchmark offloading policies 30:12 Can we make actionable insights from growing robotic sensory data? 32:17 Rationale 1: Specialization corrects errors 32:48 Model specialization can correct key errors 33:16 Rationale 2: The real world is constantly changing 34:25 Why sample?: Reduce systems costs 35:26 Minimal Images are Needed 36:20 Efficiently filter images of interest during inference 39:30 Delegate compute-intensive tasks to the cloud 40:42 Current: Multi-Robot Learning 42:20 Task-Driven Representations for Perception 43:25 Semi) Federated Learning for Robots 44:03 Control and Learning Across Data Boundaries