Event
NativeMEM improves long-horizon manipulation success rates for VLA via memory token compression
1 reports · 1 sources
Ziye Wang et al. propose NativeMEM, which reuses the VLA's own visual encoder to compress every frame of history from each camera view into a single memory token that is appended to the input sequence, enabling a pretrained VLA to attend to long-term history without an external planner or a newly initialized memory module, with minimal latency overhead. The method has been accepted to CoRL 2026. The paper reports that NativeMEM raises the success rate from 32.4% to 84.0% in simulation and up to 98.7% on a real robot, while maintaining low inference latency and GPU memory usage; using only 20% of the training data, it achieves results comparable to existing methods.
Report timeline
- NativeMEM: A Native Memory Compression Method for Long-Horizon Robotic Manipulation
NativeMEM is a vision-language-action (VLA) policy with long-horizon, real-time updated memory. Its core is the Native Memory Compression encoding scheme, which reuses the VLA's own vision encoder to compress each historical frame from each camera view into a single token. Appended to the input sequence, these tokens let the pretrained VLA attend to long-term history with negligible latency overhead, requiring neither an external planner nor a newly initialized memory module.
arXiv cs.RO ·