This repository contains the checkpoint for SimpleMemVLA, introduced in the paper SimpleMemVLA: A Simple but Effective Native-Video Memory for Vision-Language-Action Models.
SimpleMemVLA is a vision-language-action (VLA) model without a dedicated memory module. It keeps the sampled history intact and passes it to the backbone as timestamped video, using native self-attention as memory. The hidden states of a generated sub-task are the only channel to a flow-matching action head, enabling exact streaming inference with near-single-frame latency.
Links:
The model is built on the Qwen3.5-4B backbone and achieves state-of-the-art results on four memory benchmarks (RMBench, RoboMME, MIKASA-Robo, RoboMemArena) while matching the best results on general-purpose control (LIBERO). For training, evaluation, and deployment details, please refer to the GitHub repository and the paper.
This repository contains the checkpoint for SimpleMemVLA, introduced in the paper SimpleMemVLA: A Simple but Effective Native-Video Memory for Vision-Language-Action Models.
SimpleMemVLA is a vision-language-action (VLA) model without a dedicated memory module. It keeps the sampled history intact and passes it to the backbone as timestamped video, using native self-attention as memory. The hidden states of a generated sub-task are the only channel to a flow-matching action head, enabling exact streaming inference with near-single-frame latency.
Links:
The model is built on the Qwen3.5-4B backbone and achieves state-of-the-art results on four memory benchmarks (RMBench, RoboMME, MIKASA-Robo, RoboMemArena) while matching the best results on general-purpose control (LIBERO). For training, evaluation, and deployment details, please refer to the GitHub repository and the paper.
