Hello, thank you for releasing AutoHorizon.
We are evaluating AutoHorizon on RoboTwin 2.0 using the public implementation. The paper reports prediction horizons p=10 and p=50 for the LIBERO experiments, but we could not find the prediction horizon used for RoboTwin Table 3. Could you confirm the value of p used there?
We would also appreciate information about the π0.5 checkpoint used for RoboTwin, including whether it was jointly trained or task-specific and whether it is publicly available. Please also confirm the sampling step, soft-pointer implementation, and hyperparameters used for Table 3, as well as the RoboTwin configuration and evaluation seeds. If the RoboTwin evaluation code or configuration can be shared, that would be very helpful.
We currently use a jointly trained RoboTwin π0.5 checkpoint with p=32. With the public AutoHorizon implementation, the selected execution horizon is almost always 32, making its behavior close to the fixed-horizon baseline. We would like to determine whether this difference comes from the checkpoint, prediction horizon, or AutoHorizon configuration.
Thank you for your help.
Hello, thank you for releasing AutoHorizon.
We are evaluating AutoHorizon on RoboTwin 2.0 using the public implementation. The paper reports prediction horizons
p=10andp=50for the LIBERO experiments, but we could not find the prediction horizon used for RoboTwin Table 3. Could you confirm the value ofpused there?We would also appreciate information about the π0.5 checkpoint used for RoboTwin, including whether it was jointly trained or task-specific and whether it is publicly available. Please also confirm the sampling step, soft-pointer implementation, and hyperparameters used for Table 3, as well as the RoboTwin configuration and evaluation seeds. If the RoboTwin evaluation code or configuration can be shared, that would be very helpful.
We currently use a jointly trained RoboTwin π0.5 checkpoint with
p=32. With the public AutoHorizon implementation, the selected execution horizon is almost always 32, making its behavior close to the fixed-horizon baseline. We would like to determine whether this difference comes from the checkpoint, prediction horizon, or AutoHorizon configuration.Thank you for your help.