header-langage
简体中文
繁體中文
English
Tiếng Việt
한국어
日本語
ภาษาไทย
Türkçe
Scan to Download the APP
Partnered With:

Yann LeCun teaches world models to review: Is a robot's failure due to wrong action choices, or just bad luck?

Beating AI News Flash: AMI Labs, founded by Yann LeCun, together with researchers from Brown University, has published a new paper, "Counterfactual World Models," enabling the JEPA world model to infer, based on the actual outcome of one action, whether failure could have been avoided had a different action been taken at the time. This capability is called "counterfactual reasoning."

For example, a robot needs to deliver cargo to a designated location but is blown off course by a gust of wind. How can it determine whether it chose the wrong direction or whether the wind was too strong and caused the task to fail? An ordinary world model can predict what outcomes different actions typically produce, but it does not know how much impact the wind had this time.

The new method uses the deviation between the actual outcome and the original prediction to estimate the impact of this external disturbance, and then infers how the outcome would have differed had another action been chosen at the time. The entire process requires calling the existing predictor only twice, without additional training of a counterfactual model. The robot can then judge after the fact whether a different action might have had a chance to accomplish the goal.

In four physics simulation environments, the team reported that the proportion of cases in which the model retrospectively found an action that could have accomplished the goal reached 35%–69%, compared with only 14%–23% for baseline methods. Even when the baseline methods knew the true physical rules of the simulation environment, as long as they did not know the disturbance that actually occurred this time, their performance still lagged behind.

However, this method requires that the effects of actions and external disturbances be roughly additive. If a collision occurs, or a motor failure causes different commands to be affected differently, predictions may become inaccurate. Current research is still limited to simulation environments and has not yet verified actual control performance on real robots.

举报 Correction/Report
Correction/Report
Submit
Add Library
Visible to myself only
Public
Save
Choose Library
Add Library
Cancel
Finish