Stanford Seminar - Distributed Perception and Learning Between Robots and the Cloud

Overview

Save Big on Coursera Plus. 7,000+ courses at $160 off. Limited Time Only!

Grab it

Explore the challenges and opportunities of distributed perception and learning in robotics through this Stanford seminar. Delve into the complexities of robot sensory data and compute models, examining how network connectivity can enhance robotic capabilities. Investigate key challenges in cloud robotics, including distributed inference and learning. Analyze the accuracy of robot and cloud DNNs, hidden costs of network congestion, and cloud communication. Learn about dynamic decision-making in cloud offloading and the application of reinforcement learning to optimize robot-cloud interactions. Discover strategies for leveraging growing robotic sensory data, including model specialization and efficient sampling. Examine current trends in multi-robot learning, task-driven representations for perception, and federated learning for robots. Gain insights into control and learning across data boundaries in this comprehensive exploration of cutting-edge robotics research.

Syllabus

Introduction.
Robot sensory data + compute models are becoming increasingly complex.
How Can Network Connectivity Help Robots?.
Key Challenges of Cloud Robotics.
1. Distributed Inference: The Robot-Cloud Offloading Problem.
2. Distributed Learning: The Robot Sensory Sampling Problem.
Outline.
Accuracy of Robot and Cloud DNNS.
Hidden Costs of Network Congestion.
Network Costs of Cloud Communication.
Our Network Congestion Experiments.
Cloud Offloading: A Dynamic Decision-Making Problem.
Robot-Cloud Offloading: Sequential Model Selection.
Reinforcement Learning (RL).
The Robot Offloading MDP: Action Space.
The Robot Offloading MDP: State Space.
The Robot Offloading MDP: Reward.
Deep RL beats benchmark offloading policies.
Can we make actionable insights from growing robotic sensory data?.
Rationale 1: Specialization corrects errors.
Model specialization can correct key errors.
Rationale 2: The real world is constantly changing.
Why sample?: Reduce systems costs.
Minimal Images are Needed.
Efficiently filter images of interest during inference.
Delegate compute-intensive tasks to the cloud.
Current: Multi-Robot Learning.
Task-Driven Representations for Perception.
Semi) Federated Learning for Robots.
Control and Learning Across Data Boundaries.