Xiaohuan Pei*, Yuxing Chen*, Siyu Xu, Yunke Wang, Yuheng Shi, Chang Xu
Action-aware Dynamic Pruning (ADP) is a training-free, plug-and-play method for efficient VLAs. It adaptively prunes redundant visual tokens across manipulation stages by combining text-driven token relevance with an action-aware gating signal from end-effector motion—reducing FLOPs and latency while preserving task success. ADP works out-of-the-box with parallel decoding (e.g., OpenVLA-OFT).
Vision–Language–Action (VLA) models extend large vision–language models to map visual observations and language instructions into executable robot actions. In the mainstream pipeline, a vision encoder produces dense visual tokens, a projector aligns them to the language space, and an LLM fuses modalities to predict actions. However, long multi-modal sequences introduce substantial redundancy in visual tokens, increasing computation, memory usage, and latency.
Action-aware Dynamic Pruning (ADP) is a training-free, plug-and-play method for efficient VLAs. It adaptively prunes redundant visual tokens across manipulation stages by Action-aware gating signals derived from end-effector motion.
ADP significantly reduces FLOPs and latency while preserving task success rates. It works seamlessly with parallel decoding frameworks such as OpenVLA-OFT.
Our installation procedure follows the official OpenVLA-OFT repository:
Please refer to their instructions for environment setup, dependency installation, and checkpoint preparation.
If you have modified the prismatic module (or any core model implementation) before, please remember to reinstall the package in editable mode:
pip install -e .- 50–70% keep → ≤0.9% SR drop, up to 1.23× LLM speedup
- 30–40% keep → 94.4–94.8% SR, 1.29–1.35× speedup
- Spatial suite reaches 99.4% SR
- SR improves from 85.8% → 88.3%
- Latency reduces from 76.9 → 51.8 ms (1.49× speedup)
After finishing the OpenVLA-OFT installation, please replace the following folders in your OpenVLA-OFT directory with the ones provided in this repository:
experiments/prismatic/
python experiments/robot/libero/run_libero_eval_prune_v2.py \
--pretrained_checkpoint <checkpoint_path> \
--task_suite_name libero_spatial \
--qk_config_json experiments/robot/libero/configs/prune_v2_config.json(Similar commands apply to libero_object, libero_goal, libero_10.)
{
"qk_keep_enabled": true,
"qk_layer": 0,
"qk_keep_ratio": 0.75,
"qk_keep_split": [0.4, 0.6],
"qk_log_topk": 16,
"qk_debug": false,
"task_suite_name": "libero_spatial",
"use_dynamic_visual_strategy": true,
"decision_method": "adjacent",
"adjacent_variant": "extrema",
"adjacent_extrema_window": 3,
"adjacent_lookback": 2,
"initial_state": 0,
"delta_method": "net",
"L_eff": 0.15,
"min_delta_pos": 0.0,
"min_delta_rot": 0.0,
"hysteresis_up": 0.0,
"hysteresis_down": 0.0,
"tol_equal": 0.0
}-
Object Lookback Sweep (Object Suite, Window = 4–10)
🔗 Scripts |
📄 Logs -
Threshold Sweep (Pruning Ratio Sensitivity)
🔗 Scripts |
📄 Logs -
Window Keep Ablation (adjacent_lookback Sweep)
🔗 Scripts |
📄 Logs -
Spatial Lookback Sweep (Spatial Suite, Window = 4–10)
🔗 Scripts |
📄 Logs -
Main Table Ablation (Full Benchmark Comparison)
🔗 Scripts |
📄 Logs
@article{pei2025action,
title={Action-aware dynamic pruning for efficient vision-language-action manipulation},
author={Pei, Xiaohuan and Chen, Yuxing and Xu, Siyu and Wang, Yunke and Shi, Yuheng and Xu, Chang},
journal={arXiv preprint arXiv:2509.22093},
year={2025}
}We build upon:
- OpenVLA
- OpenVLA-OFT
- Hugging Face Transformers
Apache 2.0 License
