Training of RVT-2 in single tasks regime fails to generalize.

[zadaianchuk]

trafficstars

We use standard config for multi-task and train a single RL-bench task, and so far it overfits a lot, leading to 0 success rate for single tasks (e.g. slide_block_to_color_target). Is there any recommendation on how to adapt the RVT-2 training config to a single-task setup? Or is it supposed to be used only in multi-task training?

Thanks in advance for any advice.