Developing Robots that Autonomously Learn and Plan in the Real World

Explore the cutting-edge developments in autonomous robot learning and planning in this 46-minute conference talk by Glen Berseth, assistant professor at the Université de Montréal and co-director of the Robotics and Embodied AI Lab. Delve into the challenges of creating robotic agents capable of solving complex tasks and scientific problems. Learn about key concepts such as Markov Decision Processes, deep reinforcement learning, task decomposition, and lifelong learning. Discover innovative approaches like hierarchical controllers, curriculum training, and intrinsic rewards. Gain insights into representation learning for complex observations and future goals in robotics research. Understand how these advancements aim to bridge the gap between human and robotic problem-solving capabilities across various domains.

Current Challenges
Learning Agents
Markov Decision Process (MDP)
Policy Improvement
Deep Reinforcement Learning
Learning Through Interaction
Task Decomposition
Low-Level Controller (LLC)
LLC Reward
High-Level Controller (HLC)
Dynamic Obstacles
Without Hierarchy
Lifelong Learning
How to train a cleaning robot
Curriculum: Train grasping first
Curriculum: Train grasping while navigating
ReALMM: Insights
Intrinsic Rewards
SMIRL: Surprise minimization
Representation Learning for Complex Observations
Future Work: Goals
Future Work: Methods