Quick question for folks working on the Intrinsic AI for Industry Challenge.
For the cable insertion task, we have at least two distinct setup types:
-
SFP ports
-
SC ports
They differ in geometry, insertion depth, and alignment behavior.
I’m currently using a hybrid approach:
-
ACT (Action Chunking Transformer) for approach + alignment
-
Scripted FSM for insertion, force handling, retract, retry, etc.
Now that ACT is working reasonably for SFP, but, not so well for SC ports. I’m trying to decide how to structure training across these different setups.
Question
Does it make sense to:
(A) Train a single ACT policy with task conditioning (e.g., connector/port type as input)
(B) Use a shared base model with separate heads/adapters for SFP vs SC
(C) Train completely separate policies per setup
Would love to hear what others are doing here.