Robot locomotion capability with the training framework

[KZyred]

Thank you very much for providing such an excellent repository for the community.

I am curious as to why you trained in "Stage 2" with
ang_vel_yaw = [0.0, 0.0]
and heading = [0.0, 0.0]
, yet the robot in your real-world video is still able to perform rotational movements.

I look forward to your clarification on this matter.

[yinzhaodong]

Thank you for your kind words and for this thoughtful question.

In short: Setting ang_vel_yaw = [0, 0] and heading = [0, 0] in Stage 2 does not disable rotation. Stage 2 narrows the command distribution for hard parkour terrains to forward locomotion, while rotational capability is retained from Stage 1 pre-training, optionally refreshed on simple terrains when they are included in training, and can still be commanded at deployment.

1. Rotation is largely inherited from Stage 1.
   Stage 1 trains the policy with full linear and angular velocity / heading commands over diverse terrains. Stage 2 resumes from the Stage 1 checkpoint and fine-tunes on top of those weights. The policy remains command-conditioned, so at real-world deployment the operator can still provide yaw commands (e.g., via joystick), even if Stage 2 simulation training uses zero yaw on parkour terrains.

2. Zero yaw/heading in Stage 2 applies only to hard parkour terrains.
   When simple terrains are added to training, our dual command curriculum distinguishes easy and hard environments: on plane / up_stairs / down_stairs (easy terrains), the policy still samples full heading commands (inherited from the base config); on gap, stone, beam, corridor, etc. (hard terrains), we set new_ang_vel_yaw = [0, 0] and new_heading = [0, 0] to focus on forward, terrain-adaptive locomotion. Rotation is therefore preserved on flat/simple terrain, while parkour skills are trained in a forward-only manner.

We hope this clarifies the design.