Skip to content
프로젝트

SII-FUSC/AME_Locomotion has 353 GitHub stars and 41 forks. GitHot measured 0 stars in 24 hours. Last checked 2026-11-18 16:49 UTC. Last push 2026-05-05.

이 수치를 측정하는 방법

측정된 스타 증가

in 24h
0
in 7 days
기록 수집 중
in 15 days
기록 수집 중
in 1 month
기록 수집 중

서로 겹치는 기간의 순증가량이며 일별 합계가 아닙니다. 측정값이 없으면 비워 둡니다. 스타는 관심도를 나타내며 품질을 보장하지 않습니다.

소개

This repository reproduces the Attention-Based Map Encoding (AME) method from the paper Attention-Based Map Encoding for Learning Generalized Legged Locomotion.

학습

README.md

Original

AME Locomotion Reproduction (Isaac Lab + Unitree G1)

Language: 中文 | English


中文

项目简介

本项目是对论文 Attention-Based Map Encoding for Learning Generalized Legged Locomotion 中 基于注意力的地形编码器方法 AME 的复现实现。

  • 仿真与训练平台: NVIDIA Isaac Sim 5.1.0 + Isaac Lab 2.3.0
  • 机器人平台: Unitree G1 29DoF
  • 强化学习框架: RSL-RL (含本项目自定义网络扩展)

核心目标是基于高程图和注意力机制,学习具备更强地形泛化能力的腿式运动策略。

方法实现位置

AME 的主要网络实现在:

该文件包含:

  • 地形图卷积特征提取
  • 局部地形特征与本体状态的多头注意力融合
  • 与 Actor/Critic 网络对接的编码输出

安装说明

  1. 安装 Isaac Lab
  1. 安装本项目扩展
python -m pip install -e source/ame_locomotion
  1. 安装自定义 rsl_rl
python -m pip install -e rsl_rl

说明: 本项目使用了自定义 Actor-Critic 地形编码网络,若不安装本地 rsl_rl,可能无法正确导入或运行。

  1. 预训练模型目录
  • 项目内提供了预训练模型目录: pretrained/
  • 本仓库已提供预训练检查点文件(.pt),可直接用于快速测试。
  • ame1.pt: 当前默认配置训练得到(高程图 33×21,开启 CNN 下采样)。
  • ame2.pt: 在默认配置基础上加入全局上下文(attach_global=True)。

快速开始

训练:

bash run_train.sh

测试与可视化:

bash run_play.sh

两阶段训练说明

请注意 AME 采用两阶段训练流程: 第一阶段完成后,将 velocity_env_cfg_29dof.py 中的 FINETUNE 设为 True,再启动第二阶段训练。

复现说明与实现调整

整体上我们遵循论文设计,同时尝试了三点小调整。

  1. CNN 输入使用 xyz 坐标,而非仅 z 高度
  • CNN 直接处理 xyz 三维输入。
  • CNN 输出后不再额外拼接坐标。
  • 我们观察到该设置通常有更好的训练表现。
  • 可能原因是: 直接输入 xyz 能让网络更早学习到位置相关信息。
  1. 用 CNN 步长下采样降低注意力计算开销
  • 提高高程图分辨率或扩大范围时,MHA 序列长度会显著增加。
  • 我们在 CNN 阶段通过步长下采样缩短后续注意力序列。
  • 测试中未观察到明显的最终策略性能下降,同时训练开销更低。
  1. 引入 AME2 中的全局上下文
  • 我们加入了 AME2 提出的通过 MLP + max-pool 获取全局上下文信息的设计。
  • 测试发现该设计能够提升策略表现。
  • 代价是注意力权重的可解释性变差,并且训练开销明显增加。

关键文件


English

Overview

This repository reproduces the Attention-Based Map Encoding (AME) method from the paper Attention-Based Map Encoding for Learning Generalized Legged Locomotion.

  • Simulation and training stack: NVIDIA Isaac Sim 5.0.0 + Isaac Lab 2.3.0
  • Robot platform: Unitree G1 29DoF
  • RL stack: RSL-RL (with custom network extensions in this project)

The goal is to learn robust legged locomotion policies with stronger terrain generalization using elevation-map observations and attention-based terrain encoding.

Where AME Is Implemented

The core AME network implementation is in:

This file contains:

  • CNN-based terrain feature extraction
  • Multi-head attention fusion between local terrain features and proprioceptive state
  • Encoded outputs for the Actor/Critic heads

Installation

  1. Install Isaac Lab
  1. Install this extension package
python -m pip install -e source/ame_locomotion
  1. Install the local custom rsl_rl
python -m pip install -e rsl_rl

Note: This project depends on a customized Actor-Critic terrain encoder. Without installing the local rsl_rl package, imports and runtime behavior may fail.

  1. Pretrained checkpoint folder
  • A pretrained checkpoint folder is included: pretrained/
  • This repository includes released pretrained checkpoint files (.pt) for quick testing.
  • ame1.pt: trained with the current default setup (33x21 elevation map, CNN downsampling enabled).
  • ame2.pt: default setup plus global-context (attach_global=True).

Quick Start

Train:

bash run_train.sh

Play / evaluate:

bash run_play.sh

Two-Stage Training Note

Please note that AME uses a two-stage training pipeline:

  1. Finish stage-1 training first.
  2. After stage-1 is done, set FINETUNE = True in velocity_env_cfg_29dof.py.
  3. Then start stage-2 training.

Reproduction Notes and Small Deviations

Our implementation mostly follows the paper design, with three small adjustments.

  1. CNN input uses xyz coordinates instead of z-only height
  • The CNN consumes full xyz map coordinates.
  • We remove the post-CNN coordinate concatenation.
  • In our tests, this often gives better training results.
  • A likely reason is that position-related cues are learned earlier and more naturally.
  1. CNN stride-based downsampling to reduce attention cost
  • Higher map resolution or larger map coverage increases MHA sequence length.
  • We apply stride-based downsampling in the CNN stage to shorten the sequence before MHA.
  • Empirically, we did not observe clear final-policy degradation, while training cost is reduced.
  1. AME2-style global-context
  • We also add the AME2 design that extracts global context with an MLP + max-pool pathway.
  • In our tests, this improves policy performance.
  • The trade-off is weaker attention-weight interpretability and noticeably higher training cost.

Key Files