为什么值得关注

提出了一种新颖的混合学习方法，其中智能体策略基于任务目标而非参考运动进行条件化，从而能够从少量参考轨迹泛化到新的任务条件。

适合谁使用

研究人形机器人运动与操作的机器人学研究人员
从事四足/双足机器人 sim-to-real 迁移的工程师
对混合模仿学习与强化学习框架感兴趣的开发者
学习高级机器人学习技术的学生

典型使用场景

教人形机器人攀爬不同大小和位置的箱子
训练机器人将箱子移动到指定目标位置
作为目标条件化强化学习与参考方法的基准对比
通过 Web 浏览器远程评估学习策略，无需本地仿真窗口

项目优势

智能体不直接观测参考运动，从而能够泛化到新任务条件
两阶段训练（先跟踪后混合）使少量参考集也能泛化
支持无头训练与评估，利用 Viser 网页可视化实现远程访问
基于 Holosoma，利用成熟的仿真基础设施

使用前须知

非官方实现，未得到原作者认可，可能与论文方法有差异
仅在 Unitree G1 29-DoF 上验证，迁移到其他机器人需大量调整
训练需要 Isaac Sim 和特定依赖（wandb, CUDA），降低了可访问性
仅提供攀爬和搬运两种任务，未展示更广泛的任务覆盖能力

README 快速开始

Hybrid Motion Imitation

  Box Moving
  


  Box Climbing

This is an unofficial implementation and extension of GfR and HIL, built on top of Holosoma. It is not the official code release for either paper.

This release provides two tasks for the 29-DoF Unitree G1:

exp:g1-29dof-wbt-hybrid-climb
exp:g1-29dof-wbt-hybrid-object

The same hybrid learning framework is used for both box climbing and box moving. Its key ideas are:

The actor does not observe the target reference motion. We find that this design can achieve high-quality motion tracking across a wide range of tasks; see the HIL paper for more details. The actor only observes the robot state and sparse task conditions or goals, while the reference motion is used to define the goals and tracking rewards. Because the policy is conditioned on task goals rather than a reference motion, the learned behavior can easily transfer to new task conditions.
Training combines two types of environments.
- In reference-tracking environments, goals and task conditions are extracted from the reference motion and the policy is trained with tracking rewards.
- In general RL environments, goals and task conditions are randomized and the policy is trained with sparse goal-reaching rewards.
One policy is jointly optimized on two types of environments. The actor always uses the same observation interface across track and general RL envs, while the critic receives task-specific privileged information. This allows behavior learned from a small reference set to generalize to a wide range of task conditions and goals.

Setup

Run from the repository root:

bash scripts/setup_isaacsim.sh
source scripts/source_isaacsim_setup.sh
wandb login

Training

Training uses two stages:

Stage 1 trains track environments only.
Stage 2 resumes Stage 1 and trains a 50/50 mix of track and gen environments.

Climb Stage 1

CUDA_VISIBLE_DEVICES=0 \
python src/holosoma/holosoma/train_agent.py \
  exp:g1-29dof-wbt-hybrid-climb \
  logger:wandb \
  --training.num-envs=8192 \
  --algo.config.num-learning-iterations=15000 \
  --simul

项目描述

Unofficial implementation and extension of GfR (RSS 2026) and HIL (TOG 2026)

Hybrid-Motion-Imitation