BHR Lab: BIT Humanoid Robot Isaac Lab

Just in test.

Temporary record

Example: BHR8 FC2 without arm joints

Tasks definition

Task Name	Terrain	Randomization	Self Collision	Mirror
bhr8_fc2_noarm_flat_nsc	flat	off	off	off
bhr8_fc2_noarm_flat	flat	off	on	off
bhr8_fc2_noarm_flat_random	flat	on	on	off
bhr8_fc2_noarm_rough_random	rough	on	on	off
bhr8_fc2_noarm_flat_nsc_mirror	flat	off	off	on
bhr8_fc2_noarm_flat_mirror	flat	off	on	on
bhr8_fc2_noarm_flat_random_mirror	flat	on	on	on
bhr8_fc2_noarm_rough_random_mirror	rough	on	on	on

Training procedure (on RTX5080)

First training:

python scripts/rsl_rl/train.py --task=bhr8_fc2_noarm_flat_nsc_mirror --headless --num_envs=10240 --max_iterations=3000

Second training:

python scripts/rsl_rl/train.py --task=bhr8_fc2_noarm_flat_mirror --headless --num_envs=10240 --max_iterations=3000 --resume

Third training:

python scripts/rsl_rl/train.py --task=bhr8_fc2_noarm_flat_random_mirror --headless --num_envs=10240 --max_iterations=3000 --resume

Fourth training:

python scripts/rsl_rl/train.py --task=bhr8_fc2_noarm_rough_random_mirror --headless --num_envs=4096 --max_iterations=6000 --resume

vscode settings:

ctrl + shift + p
tasks: run task
setup_python_env

Create Lite Lab

To create a lightweight version of BHR Lab with essential files only:

python scripts/create_lite_lab.py --path /absolute/path/to/target --name project_name

This script will:

1. Create necessary directories (.vscode, scripts/rsl_rl, envs)
2. Copy VSCode settings
3. Copy task initialization files
4. Copy and modify training scripts with correct import paths

Troubleshooting

2025-05-27

• Issue: After updating to the latest version of Isaac Lab, the system prompts missing numba package
• Solution: Install numba using pip

pip install numba

2025-06-07

• Issue: IsaacLab API breaking change - quat_rotate_inverse deprecated (Issue #2129)
• Problem:
  • IsaacLab #2129 changed quat_rotate_inverse to quat_apply_inverse
  • This causes dimension errors in lateral_distance reward function
  • Old code produces warnings and potential calculation errors
• Root Cause: API function name and signature changes in latest IsaacLab versions
• Solution:
  1. Update IsaacLab to the latest version
  2. Replace quat_rotate_inverse with quat_apply_inverse in reward functions
  3. Verify dimension compatibility in affected reward calculations
• Files Affected:
  • Reward functions using quaternion rotations
  • Specifically lateral_distance reward computation
• Note: Older IsaacLab versions may not have quat_apply_inverse function, update is required