Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..

VLA-Cache: Efficient Vision-Language-Action Manipulation via Adaptive Token Caching

Authors: Siyu Xu, Yunke Wang, Chenghao Xia, Dihao Zhu, Tao Huang, Chang Xu

NeurIPS 2025 | Venue PDF | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive experiments on two simulation platforms (LIBERO and SIMPLER) and a real-world robotic system demonstrate that VLA-Cache achieves up to 1.7 speedup in CUDA latency and a 15% increase in control frequency, with negligible loss on task success rate.
Researcher Affiliation Academia 1School of Computer Science, University of Sydney, Australia 2John Hopcropt Center for Computer Science, Shanghai Jiao Tong University, China
Pseudocode Yes Algorithm 1 Dynamic Token Selection
Open Source Code Yes The code and videos can be found at our project page: https://vla-cache.github.io.
Open Datasets Yes We evaluate VLA-Cache on robotic manipulation tasks across two simulated environments (LIBERO [17] and SIMPLER [18])
Dataset Splits Yes We collected 218 valid demonstrations for the training dataset, with slight random adjustments to the initial positions of the pot and the robotic arm in each episode. During evaluation, each trial is recorded as a success (1) or failure (0); there is no partial credit. ... We run 20 trials for Pick Pot and Put Sausage, and 30 trials for Place Cube and Wipe Table, for a total of 100 trials per method.
Hardware Specification Yes All experiments are conducted on an NVIDIA RTX 4090 GPU.
Software Dependencies No Simulated evaluations for Cog ACT and SIMPLER are conducted on a single NVIDIA RTX 4090 GPU in BF16 precision. During inference, we use DDIM sampling with 10 steps and a classifier-free guidance (CFG) coefficient of 1.5. Similarly, for Open VLA and LIBERO, inference is performed on a single NVIDIA RTX 4090 GPU in BF16 precision. ... Training on the real robot used Lo RA-based fine-tuning for 50,000 steps... We employed Py Bullet as an inverse kinematics (IK) controller. The paper mentions software components and techniques but lacks specific version numbers for key libraries or frameworks needed for replication.
Experiment Setup Yes Unless specified otherwise, we use a static token similarity threshold Ο„ = 0.996, top-k = 100 for retained static tokens, and a task-relevance threshold Ο„task = 0.5. These parameters are applied consistently across all simulated and real-world settings, including SIMPLER with Cog Act. For real-world Jaco2 experiments, we slightly reduce the similarity threshold to Ο„ = 0.85 to accommodate environmental noise. Training on the real robot used Lo RA-based fine-tuning for 50,000 steps