AI for Manipulation Paper Presentation

How would you describe dexterous manipulation?

What are some things that you found interesting about either of the papers?

How are the two approaches to learning dexterous manipulation similar? (Select all that apply)

• Both approaches use the Shadow Dexterous Hand

• They both use point clouds in order to map the environment

• Both approaches conduct simulations for training

• The goal of both approaches was to better generalize to novel goals

How do the two approaches differ? (Select all that apply)

• DexDeform uses human demonstration

• OpenAI uses reinforcement learning, DexForm doesn't

• DexDeform deals with deformable objects, OpenAI doesn't

• DexDeform uses reinforcement learning, OpenAI doesn't

What are the three steps of the DexDeform?

• 1. 1. learning from videos, 2. skill abstraction, 3. exploration with novel goals

• 1. 1. learning from manually moving the robot, 2. skill abstraction, 3. exploration with novel goals

• 1. learning via teleoperation, 2. manually labeling the actions, 3. exploration with novel goals

• 1. 1. learning via teleoperation, 2. skill abstraction, 3. exploration with novel goals

In the OpenAI paper, the policy is trained with Proximal Policy Optimization (PPO), which requires a policy network and a _____ network

Do you have any questions about the paper that our presentation didn't cover?