OpenVLA

“On the LIBERO benchmark, removing half of the LLM blocks actually *improved* OpenVLA-OFT's success rate from 95.0% to 98.3% under the same fine-tuning budget”

“OpenVLA (OpenVLA-OFT): An open-source VLA model built on a Llama-2-7B backbone.”

“Patch Policy "surpasses fine-tuned OpenVLA-OFT by 18% while using roughly 0.7% of the parameters"”

“S2-VLA consistently outperforms larger 7B-scale models like OpenVLA-OFT and MemoryVLA”

“"Temporal smoothing trades success for stability on every single-step backbone: on OpenVLA it raises aggregate stability from 20.1% to 36.2% but reduces mean success from 36% to 28%"”

“"a single-step OpenVLA policy attains a success rate of only 36%, while its chunked, memory-augmented variant OpenVLA-OFT reaches 92%"”

“the architecture can be layered onto existing investments (e.g., OpenVLA, π0, or other foundation models) rather than requiring a full model replacement.”

“StarVLA-OFT attaches a lightweight MLP regression head... following OpenVLA-OFT”