VLA-ACL: patch scorers for visual token pruning in OpenVLA-OFT

VLA-ACL (Action Consistency Learning) learns which vision tokens a Vision-Language-Action model needs. A lightweight patch scorer is trained on top of a frozen, fine-tuned OpenVLA-OFT policy so that the actions predicted from the top-K vision patches per camera view stay consistent with the full-context (teacher) actions and with the ground-truth demonstration actions. At inference time only the top-K patches of each view are passed to the LLM backbone.

This repository contains 8 scorer checkpoints, one for each LIBERO task suite at two pruning budgets:

K (patches kept per view)	patches kept (2 views, 256 each)	pruning rate
32	64 of 512	87.5%
64	128 of 512	75%

Files

scorer-libero-spatial-k32.pt   scorer-libero-spatial-k64.pt
scorer-libero-object-k32.pt    scorer-libero-object-k64.pt
scorer-libero-goal-k32.pt      scorer-libero-goal-k64.pt
scorer-libero-10-k32.pt        scorer-libero-10-k64.pt

The file name is scorer-libero-{suite}-k{K}.pt. Each file is a state_dict (about 130 MB, bf16, 65.2M parameters) of the patch scorer: a 5-layer bidirectional Transformer (hidden size 1024, 16 heads, MLP ratio 4, pre-norm, GELU, no dropout) with a linear input projection from the 2176-dimensional OpenVLA-OFT vision embeddings and a linear logit head. Parameter names: input_proj.*, encoder.*, logit_head.*.

A scorer only works with the OpenVLA-OFT checkpoint of the matching suite it was trained against (moojink/openvla-7b-oft-finetuned-libero-{spatial,object,goal,10}).

Evaluation

LIBERO success rate (%), 50 trials per task (500 episodes per suite), averaged across 3 random seeds, K patches kept per view:

	Spatial	Object	Goal	Long	Average
K=32	98.2	98.2	96.6	88.7	95.4
K=64	98.7	98.4	96.5	94.2	97.0

Citation

@misc{du2026vlaaclactionconsistentvisualtoken,
      title={VLA-ACL: Action-Consistent Visual Token Pruning for Efficient Vision-Language-Action Models}, 
      author={Owen Du and Yang Yue and Jie Zhang and Jiaqi Pi and Chi Bene Chen and Gao Huang},
      year={2026},
      eprint={2610.08133},
      archivePrefix={arXiv},
      primaryClass={cs.RO},
      url={https://arxiv.org/abs/2610.08133}, 
}