LIBERO state-to-action policy
A feed-forward policy from an 8-dimensional robot state to a 7-dimensional action. Weights were trained from scratch on a slice of HuggingFaceVLA/libero.
Size
- Parameters: 50,409,479
- Checkpoint file: 201,641,637 bytes (about 201.6 MB of fp32 weights)
Data and split
- Source:
HuggingFaceVLA/libero, train split, first 4,000 frames - Training episodes: 12
- Held-out episodes: 3
Local score
Offline episode success: 1 / 3 (0.3333).
An episode counts as a success when the policy's mean squared action error on that episode is at or below the median per-frame error on the training frames. The threshold was 0.1453.
This is not a simulator rollout. The LIBERO leaderboard number is the success rate of rollouts in simulation. That number has not been measured for this checkpoint.
Reproduce
train.py rebuilds the same network. policy.py loads model.pt and maps a length-8 state to a length-7 action.