带有命名地点的始终在线二维激光雷达SLAM，作为大型语言模型代理的空间记忆。

英文 |中文

---

你好，世界

Pixi run -e ros 演示：制图师的公共revo_lds包（Neato Revo LDS，16分钟 在浏览器调试视图中以8倍速度重播。会话#0和#1依次冻结，#2开始。 绘制地图。

一位被要求从冰箱取可乐的代理理解请求，但不理解家庭：其中 冰箱是通往厨房的方向，可乐最后被看到的地方。无论用什么型号，它 需要空间记忆：它必须能够记住某个地点，无论是命令、一句话还是一张照片， 稍后让机器人回去，并在上下文压缩和映射中保留内存 重新优化和流程重启。EvergreenSLAM 是一款围绕此设计的二维激光雷达 SLAM 需求。

• 一个模式，无操作符。 它作为后台进程运行，没有单独映射，且 本地化模式：它总是映射到当前的会话，同时本地化对会话 它已经冻结了，会话在添加足够多的新区域后会冻结自己，而且它的 体式受限。
• **按区域划分，不按运行时间划分。

EvergreenSLAM

Always-on 2D lidar SLAM with named places, as a spatial memory for LLM agents.

English | 中文

---

https://github.com/user-attachments/assets/ecc5e8f2-d290-4b70-8356-79e4789ae43c

An agent (Claude Code) using EvergreenSLAM as the memory of a home robot; 1:44, wall clock and playback speed on screen. Day 1, from an empty map: asked for the balcony, it finds it through one line of its own exploration notes. Day 2, after a night powered off: a new session with no chat history finds the scissors from memory. No teleoperation, no mapping or localization by hand. Full video (3:18): YouTube; details under Experiments.

Hello, world

$ egs here
current places/dock 2.4m
closest places/dock 2.4m frozen; places/kitchen 15.3m frozen
robot 0.48,1.02,-2.72  keyframe 0s  @solve 135
base 0.75 (cover 98%)  since base match 0.0m  match 0.70 (avg 0.70)

An agent asked to fetch a cola from the fridge understands the request but not the home: where the fridge is, how to reach the kitchen, where the cola was last seen. Whatever model is used, it needs a spatial memory: it must be able to remember a place from a command, a sentence or a photo, send the robot back there later, and keep that memory across context compaction, map re-optimization and process restarts. EvergreenSLAM is a 2D lidar SLAM designed around this requirement.

• One mode, no operator. It runs as a background process with no separate mapping and localization modes: it always maps into the current session while localizing against sessions it has already frozen, and a session freezes itself once it has added enough new area and its poses are well constrained.
• Bounded by area, not uptime. Trimming marginalizes what it removes and sparsifies the result to a Chow-Liu tree, and frozen poses are held constant in the optimizer, so that memory and CPU are bounded by the mapped area rather than by uptime.
• Places that keep resolving. A place is an anchor bound to a submap of the pose graph rather than to coordinates, exposed to the agent as a directory tree plus a small CLI (egs, also served as MCP tools).
• One cheap sensor. A 2D lidar and a CPU; no IMU, wheel odometry, camera or GPU. It has been run on a 215 CNY lidar and an RK3588 board, and on the public 2D datasets.

The scan matching and the submaps follow Cartographer; the lifelong layer above them is the project’s own. How it relates to slam_toolbox is under Related work.