Repository navigation
Go2 RL Gym → Go2 RL Lab
go2_rl_gym is now go2_rl_lab, built on Isaac Lab 3.0 EA / Isaac Sim 6.1, with Newton MJWarp as the physics backend
The original Isaac Gym implementation is preserved on the gym branch; the current Isaac Lab implementation uses the lab branch
Policy performance
Single-stage training and single-policy deployment, using only proprioceptive sensing at deployment
- Terrain traversal: 20 cm stairs, 30-degree slopes, and 30 cm obstacles
- Flat-ground locomotion: up to 4 m/s in both forward and backward directions
- Best RoboGauge score: 0.6715 → 0.6929
- Maximum angular velocity command: 2 → 3 rad/s
Training performance
RTX 4090 · MoE-CTS · Visualization disabled · 50 measured iterations after 10 warm-up iterations
| Parallel environments | Framework | Steady-state GPU memory (GiB) | Rollout time (s/iter) | Total iteration time (s/iter) |
|---|---|---|---|---|
| 8192 | Gym | 9.76 | 1.78 | 2.30 |
| 8192 | Lab | 9.13 | 1.72 | 2.65 |
| 16384 | Gym | 15.63 | 3.79 | 4.70 |
| 16384 | Lab | 17.80 | 2.68 | 4.37 |
Algorithm and environment
- Parallelism and history: increase the default environment count from 8192 to 16384 and observation history from 5 to 10 frames
- Heights and augmentation: use RayCaster for terrain-height queries and introduce left-right symmetry augmentation, absent from the Gym version
- Motors and terrains: use Unitree's official Go2HV motor model as a reference; add tyre and small-column terrains
- Terrain-dependent limits: dynamically adjust velocity-command limits according to the terrain at the robot's current position
Engineering support
The unified go2_rl_lab.sh entry point supports distributed training, recovery of invalid environments, reward clipping, and SwanLab / W&B logging. All Sim2Sim / Sim2Real code has been adapted
go2_rl_gym 正式更名为 go2_rl_lab,基于 Isaac Lab 3.0 EA / Isaac Sim 6.1 构建,物理后端切换为 Newton MJWarp
旧版 Isaac Gym 实现保留在 gym 分支,当前 Isaac Lab 实现使用 lab 分支
策略表现
单阶段训练、单策略部署,部署仅使用本体感知
- 多地形越障:20 cm 楼梯、30 度斜坡、30 cm 障碍
- 平地运动:前进、后退速度最高均为 4 m/s
- RoboGauge 最优得分:0.6715 → 0.6929
- 最大角速度指令:2 → 3 rad/s
训练性能
RTX 4090 · MoE-CTS · 关闭可视化 · 预热 10 个迭代后统计 50 个迭代
| 并行环境数 | 框架 | 稳态显存(GiB) | 采样时间(s/iter) | 总迭代时间(s/iter) |
|---|---|---|---|---|
| 8192 | Gym | 9.76 | 1.78 | 2.30 |
| 8192 | Lab | 9.13 | 1.72 | 2.65 |
| 16384 | Gym | 15.63 | 3.79 | 4.70 |
| 16384 | Lab | 17.80 | 2.68 | 4.37 |
算法与环境
- 并行与历史:默认并行环境数 8192 → 16384,观测历史 5 → 10 帧
- 高程与增强:使用 RayCaster 计算地形高程,引入左右对称数据增强(Gym 中没有该增强)
- 电机与地形:参考宇树官方 Go2HV 电机模型,新增轮胎和小方柱地形
- 地形限速:根据机器人当前位置所在的地形,动态调整速度指令上限
工程支持
统一入口 go2_rl_lab.sh,支持分布式训练、异常环境恢复、奖励异常裁剪以及 SwanLab / W&B 日志,已完成所有 Sim2Sim / Sim2Real 代码适配