[One-page summary] Active Task Randomization: Learning Visuomotor Skills for Sequential Manipulation by Proposing Feasible and Novel Tasks (CVPR 2023) by Pang et al.

● Summary: Active Task Randomization can create tasks for training the skill policies to handle diverse scenarios in sequential manipulation tasks

 

● Approach highlight

• Active task Randomization: task sampler propose suitable tasks in the simulation

• Task Encoder: adaptively estimate the feasibility of sampled tasks

 Maps each task parameter into a compact embedding($\rightarrow$○) in the replay buffer and estimate the novelty

 

●  Main Results

●   Discussion

• is it a fair comparison to the baseline?