GDN-Memory：具身世界模型的恒定占用联想记忆系统
基于 Gated DeltaNet 快权重构建的三层（物理/物体/行为）恒定占用记忆系统， 集成于 Genie Envisioner 的 Slow-Fast 推理流水线。总占用 256 KB，更新延迟 <1ms，联想检索精度 >85%。
GDN-Memory: Constant-Footprint Embodied Memory for World Models
Gated DeltaNet hierarchical memory integrated into Genie Envisioner's Slow-Fast VLA pipeline. 256KB constant footprint, <1ms update latency, targeting ICRA/CoRL 2026.
•
1 min read
GDN-Memory: Constant-Footprint Embodied Memory for World Models
Gated DeltaNet hierarchical memory integrated into Genie Envisioner's Slow-Fast VLA pipeline. 256KB constant footprint, <1ms update latency, targeting ICRA/CoRL 2026.
•
2 min read
GDN-Memory: Constant-Footprint Associative Memory for Embodied World Models
A three-layer (physical / object / behavioral) constant-memory system built on Gated DeltaNet fast weights, integrated into Genie Envisioner's Slow-Fast inference pipeline. 256 KB …
Efficient Inference Infrastructure for Deployable Vision-Language-Action Policies
Inference optimization stack for flow-matching VLA policies achieving sub-20ms latency on commodity hardware via speculative denoising and pipeline scheduling.
Open-Source Robotics Collaboration with 子豪兄
Joint robotics projects with Zihao (同济大学 B站 UP主) — OpenDuckMini bipedal robot, quadruped platforms, and community hardware documentation co-authored at Tongji University.
•
1 min read
