DRIVE LIBERO Policy

This repository contains the exported DRIVE LIBERO policy checkpoint used for the main simulation result (the nft2 checkpoint in the anonymous experiments). It is based on allenai/MolmoAct2-LIBERO and predicts continuous 7-DoF end-effector action chunks with horizon 10 from agent-view and wrist-view observations plus the language instruction.

Reported LIBERO evaluation

The same policy checkpoint was evaluated in three independent runs. Values are mean success rate (%) plus or minus standard deviation.

Suite Success rate
LIBERO-Spatial 99.00 +/- 1.00
LIBERO-Object 99.67 +/- 0.58
LIBERO-Goal 98.00 +/- 1.00
LIBERO-Long 96.33 +/- 1.53
Average 98.25 +/- 0.43

Files

  • model.safetensors: policy weights.
  • config.json: policy and model configuration.
  • policy_preprocessor.json and associated tensors: input preprocessing.
  • policy_postprocessor.json and associated tensors: action postprocessing.
  • nft_export_manifest.json: sanitized export metadata.

Training-data note

The full DRIVE training mixture contains expert demonstrations, autonomous successful rollouts, autonomous failed rollouts, and human-in-the-loop data. These sources share the same trajectory schema. The initial anonymous code release uses the original LIBERO-format data to test the complete pipeline; additional collected trajectories are not included in this model repository.

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.

Downloads last month
9
Safetensors
Model size
5B params
Tensor type
F32
·
BF16
·
Video Preview
loading