Proper way to train wbt with terrain

[githubLeoliu]

Does anyone know what is the proper command to train wbt with terrain? I ran the command under holosoma/config_values/webg/g1/experiment.py
python src/holosoma/holosoma/train_agent.py \ exp:g1-29dof-wbt \ terrain:terrain-load-obj \ --terrain.terrain-term.obj-file-path="src/holosoma/holosoma/data/motions/g1_29dof/whole_body_tracking/terrain_slope.obj" \ --command.setup_terms.motion_command.params.motion_config.motion_file\ ="src/holosoma/holosoma/data/motions/g1_29dof/whole_body_tracking/motion_crawl_slope.npz" \ --simulator.config.scene.env_spacing=5.0 \ --training.headless=False
I notice that the robot does not align with the terrain. This misalignment does not happen when training with interactive object.

Also, the code only support for .obj terrain. The motion files in omni-retarget dataset are align with .urdf file. Are additional conversion from .urdf to .obj needed for properly train wbt with terrain?

[Juyue]

1. RE: robot does not align with the terrain:
   Try replacing simulator.config.scene.env_spacing=5.0 with simulator.config.scene.env_spacing=0.0?
2. RE: This misalignment does not happen when training with interactive object..
   Since env_spacing is set to 0.0 in g1_29dof_wbt_w_object
3. RE: The motion files in omni-retarget dataset are align with .urdf file. Are additional conversion from .urdf to .obj needed for properly train wbt with terrain?
   Yes, one needs to properly transfer .urdf to .obj to train with existing wbt training code.

[githubLeoliu]

1. When setting simulator.config.scene.env_spacing=0.0all robots are initialized on the same terrain, while the other terrains remain empty. Is this the expected behavior?
   My understanding was that, similar to g1_29dof_wbt_w_object, each environment should have its own terrain instance, so that different robots act on different terrains. With env_spacing = 0.0, I also observe some robots being initialized inside or passing through terrain boundaries (e.g., slopes or walls).

python src/holosoma/holosoma/train_agent.py \
  exp:g1-29dof-wbt-w-object \
  --training.num_envs=10 \
  --simulator.config.scene.env_spacing=2.0 \
  --training.headless=False

2. When setting simulator.config.scene.env_spacing=2.0, object interaction works as expected.From the code, this is handled in holosoma/simulator/isaacsim/registry_utils.py register_individual_objects which clearly registers per-environment object instances. However, I do not observe a similar per-environment registration logic for terrain interaction. It seems that terrain is shared globally rather than instantiated per environment, unlike objects.

python src/holosoma/holosoma/train_agent.py \
  exp:g1-29dof-wbt \
  terrain:terrain-load-obj \
  --training.num_envs=10 \
  --simulator.config.scene.env_spacing=0.0 \
  --terrain.terrain-term.obj-file-path="src/holosoma/holosoma/data/motions/g1_29dof/whole_body_tracking/terrain_slope.obj" \
  --command.setup_terms.motion_command.params.motion_config.motion_file\
="src/holosoma/holosoma/data/motions/g1_29dof/whole_body_tracking/motion_crawl_slope.npz" \
  --training.headless=False

3. For objects, holosoma/config_values/wbt/g1/experiments.py specifies the object via a URDF file, e.g. objects_largebox.urdf. This provides a clear entry point for defining object properties such as collision, scale, and physical parameters. In contrast, terrain does not seem to have a similar URDF-based entry point. Instead, terrain geometry must be provided directly as a .obj file. Is this expected?