Ev Zisselman, Mirco Mutti, Shelly Francis-Meretzki, Elisei Shafer, Aviv Tamar
Best Paper Award at ICML EXAIT
NeurIPS 2025
Note: We record observations without the mask for the BC algorithm!
Cloning blindfolded experts achieves better generalization compared with the cloning of standard experts in all tasks.
Failure
Success
Cloning the blindfolded expert leads to more exploratory behavior that generalizes better to test levels.
Note that even for failure cases, cloning the blindfolded expert still explores the maze.
Failure - limited exploration
Failure - but still explors
Examples of the expert's views during demonstrations:
The full observations (Expert) vs. the masked observations (Blindfolded Expert).
Each observation comprises a frame from wrist camera 1 (left) and wrist camera 2 (right).
Note: We record the observations without the mask for the BC algorithm!
wrist 1 wrist 2
Cloning blindfolded experts achieves better generalization compared with the cloning of standard experts in all tasks.
The robot learns task-dependent behavior for training:
Aligning the shapes before inserting.
The robot learns a general behavior for training:
Exploring the domain and searching for the insertion angle and position.
Fails to align the shape when handling previously unseen pegs (test).
Explores the domain and finds the insertion pose of all test shapes (test).