AutoPruner OFT Object: four retained decoder layers

This physical checkpoint removes original zero-based decoder layers 2–29 from OpenVLA-OFT Object and retains layers 0, 1, 30 and 31. The selected recovery saves step 10000, an MLP boundary adapter with bottleneck 1024, action head and proprio projector. The source is the artifact path used by the paper's Object evaluation.

Paper · Code and installation. The paper reports 463/500 successes. That historical score is not a fresh reproduction. Preparation on 2026-10-30 verified two native action inputs against the source recovered policy with maximum error 0.0; ten Object tasks with one standard initial-state trial each completed at 9/10 under OSMesa. Those ten trials are a bounded check and are not an independent final benchmark or a replacement for the paper's 500 episodes.

Use

Follow the installation guide to install the pinned OFT environment and complete LIBERO simulation assets. Download all files from this repository together, then run:

python -m autoprune.eval_oft_export --exported /path/to/this/model \
  --suite libero_object --trials 1 --max-eval-tasks 1 --run-root runs

No original search directory, dense reference, training data, teacher cache or LLM API is needed for this evaluation. Do not prune this checkpoint a second time. The command reads deletion indices and recovery components from autoprune_export.json. Keep all tokenizer, modeling, normalization and recovery files together. This checkpoint is for LIBERO Object simulation; no real-robot or cross-embodiment validation is claimed.

Sources and notices

Base model: moojink/openvla-7b-oft-finetuned-libero-object. The upstream model card currently labels its license MIT. Retain applicable base model and transitive component terms as well as the accompanying upstream and framework notices. The code license does not relicense unrelated external assets.