G1 Fast Running¶

Purpose¶

Push the working running policy to a higher forward velocity curriculum while keeping the running reward settings mostly unchanged.

Item	Value
Task ID	`Unitree-G1-29dof-Running-Fast`
Config	`source/unitree_rl_lab/unitree_rl_lab/tasks/locomotion/robots/g1/29dof/fast_running_env_cfg.py`
Parent task	`Unitree-G1-29dof-Running`
Output experiment root	`logs/rsl_rl/unitree_g1_29dof_running_fast/`
Warm-start checkpoint	`logs/rsl_rl/unitree_g1_29dof_running/2026-05-14_03-36-49_running_finish_may14/model_2100.pt`

fast_running_env_cfg.py subclasses RunningRobotEnvCfg, then changes only the command ranges. It intentionally keeps the gait, clearance, energy, action-rate, and termination settings from the normal running task.

Tweaks From Running¶

Area	Running task	Fast running task
Initial forward command	`lin_vel_x=(0.2, 0.8)`	`lin_vel_x=(0.8, 1.6)`
Initial lateral command	`lin_vel_y=(-0.05, 0.05)`	`lin_vel_y=(-0.05, 0.05)`
Initial yaw command	`ang_vel_z=(-0.1, 0.1)`	`ang_vel_z=(-0.1, 0.1)`
Forward curriculum limit	`lin_vel_x=(0.0, 2.0)`	`lin_vel_x=(0.0, 3.0)`
Lateral curriculum limit	`lin_vel_y=(-0.2, 0.2)`	`lin_vel_y=(-0.15, 0.15)`
Yaw curriculum limit	`ang_vel_z=(-0.3, 0.3)`	`ang_vel_z=(-0.25, 0.25)`

The first question was whether the learned running gait could stretch to higher forward speeds before adding more reward changes.

Runs¶

The fast-running task was warm-started by copying the normal running checkpoint into the fast experiment root:

logs/rsl_rl/unitree_g1_29dof_running_fast/from_running_model_2100/model_2100.pt

Fast training was started with:

TERM=xterm /home/zeul/IsaacLab/isaaclab.sh -p scripts/rsl_rl/train.py \
  --headless \
  --task Unitree-G1-29dof-Running-Fast \
  --resume \
  --load_run from_running_model_2100 \
  --checkpoint model_2100.pt \
  --run_name fast_from_running_2100 \
  --max_iterations 3000

The fast run writes to:

logs/rsl_rl/unitree_g1_29dof_running_fast/2026-05-14_03-57-12_fast_from_running_2100/

The completed fast model reached:

logs/rsl_rl/unitree_g1_29dof_running_fast/2026-05-14_03-57-12_fast_from_running_2100/model_5099.pt

Playback exported:

logs/rsl_rl/unitree_g1_29dof_running_fast/2026-05-14_03-57-12_fast_from_running_2100/exported/policy.pt
logs/rsl_rl/unitree_g1_29dof_running_fast/2026-05-14_03-57-12_fast_from_running_2100/exported/policy.onnx

Observed Behavior¶

The completed fast run reached the full 3.0 m/s curriculum. Final TensorBoard values around model_5099.pt were:

Metric	Approximate final value
`Train/mean_reward`	`29.6`
`Train/mean_episode_length`	`986 / 1000`
`Metrics/base_velocity/error_vel_xy`	`0.49`
`Metrics/base_velocity/error_vel_yaw`	`2.33`
`Curriculum/lin_vel_cmd_levels`	`3.0 m/s`
`Episode_Termination/bad_orientation`	about `2%`
`Episode_Termination/time_out`	about `93%`

GUI playback was visually interesting and stable enough to justify a 10 m/s stress-test variant.

The first playback used the inherited small play terrain, which can let robots run off the edge at high speed. This is a playback artifact risk: training used a larger terrain than the default play config. The later sprint task explicitly enlarges the play terrain.