Spring 2022 GRASP SFI - Ted Xiao, Robotics at Google, “A Panorama of End to end Robot Learning”

In this talk

1. 1:55

   Why expand end-to-end learning

   Xiao compares robotics with fields such as speech and computer vision, where learning has grown from individual pipeline components to larger systems trained end to end. He argues that robotics could follow a similar path, while acknowledging that the field faces distinct challenges.

2. 6:36

   Where learning fits in robotics

   He describes a spectrum from classical control in stable settings to systems that combine perception and control in a neural network. Factories can benefit from carefully tuned, predictable methods, while changing warehouses and unfamiliar environments create a stronger need for learning and adaptation.

3. 13:02

   Reasons to pursue end-to-end methods

   Xiao points to robot learning results including QT-Opt, which he says generalized grasping to more than 99% of unseen objects, alongside multitask and imitation learning systems. He also cites examples of perception and action working together in bees and athletes, and argues that detailed rules struggle to capture the complexity of everyday tasks. He connects this view to Richard Sutton’s argument that methods benefiting from more data and compute can improve over time.

4. 25:32

   Data, training, and system efficiency

   Xiao discusses learning from broad play data, using hindsight relabeling to create training goals, and applying conservative Q-learning to avoid treating every observed action as a success. He describes sim-to-real image translation and concurrent control, which he says increased execution speed by almost 50%. He also presents AW-Opt as a way to combine imitation learning with reinforcement learning, and stresses that exploration data should remain useful for later training.

5. 36:04

   Building toward broader robot capabilities

   Xiao sees potential in combining robot datasets with language, vision, and human activity data, while noting that datasets from different labs can be difficult to clean and align. He suggests using semantic knowledge from varied sources as a transferable aid while robotics scales up its own data and methods. In the discussion, he calls human activity data a possible middle ground between internet content and robot-specific datasets.