DRIVE LIBERO Critic
This repository contains the action-chunk critic used for the DRIVE LIBERO experiments. Given a visual-language condition and a candidate action chunk, the critic predicts a scalar score used to rank candidate chunks.
Architecture and training configuration
- Condition encoder: allenai/MolmoAct2-LIBERO
- Action horizon: 10
- Action dimension: 7
- Critic hidden dimension: 512
- Action embedding dimension: 256
- Dropout: 0.1
- Discount factor: 0.95
- Target-network EMA coefficient: 0.005
- Learning rate: 5e-5
- Training steps: 20,000
The checkpoint is intended to be loaded with the DRIVE release code. It is not a standalone control policy.
Files
- critic_last.pt: critic weights, target critic weights, optimizer state, and sanitized training configuration.
- critic_config.json: minimal architecture and scoring configuration.
Anonymity
This artifact contains no author identity, experiment-tracking identifier, or machine-specific dataset/output path. The repository is private during review and can be made public when the anonymous release is ready.