A Distributional View on Multi-Objective Policy Optimization

In the videos below, the color of the object and of the robot fingertips indicates the amount of impact force. Red indicates a high impact force (of 5N), dark orange indicates medium, and yellow indicates low. When the goal is reached, the object turns green and the episode terminates.

Touch

goal: touch block with >5N force

scalarized MPO

1: does task quickly, but pain is high

2: does task slowly, with low pain

3: no pain, but doesn't solve task

MO-MPO (ours)

4: does task quickly, with low pain

5: does task quickly, with low pain

6: does task with minimal pain

Turn

goal: turn dial to target angle (blue)

scalarized MPO

1: good at task, but high pain penalty

2: low pain

3: very low pain, but struggles at task

MO-MPO (ours)

4: as good at task as 1, with less pain

5: as low pain as 2, and better at task

6: as low pain as 3, and better at task

Page updated

Google Sites

Report abuse