Error about using aloha_split robot

[jbpen]

@Axi404
Hello, I refer to the tutorial to generate dual-arm data，the minimal.yaml as follows, I just change robot and didn't change the code or other parts. I encountered an error when I was executing the demogen.py.

error:
2025-12-15 17:18:54,762 - genmanip.utils.standalone.utils - ERROR - motion planning failed at step: 3
2025-12-15 09:18:54 [43,437ms] [Error] [genmanip.utils.standalone.utils] motion planning failed at step: 3
2025-12-15 17:18:54,763 - genmanip.utils.standalone.utils - ERROR - Task not completed, retry......
2025-12-15 09:18:54 [43,437ms] [Error] [genmanip.utils.standalone.utils] Task not completed, retry......

demonstration_configs:
- domain_randomization: &id001
    cameras:
      config_path: configs/cameras/fixed_camera_aloha_refined.yml
      type: fixed
    random_environment:
      has_wall: false
      hdr: false
      robot_base_position: false
      robot_eepose: false
      table_texture: false
      table_type: false
      wall_texture: false
    rewrite_instruction: false
  generation_config: &id002
    action_path:
      mode: auto
      robot: 0
    articulation: []
    goal:
    - - obj1_uid:
        - '0'
        obj2_uid:
        - '1'
        position:
        - top
        arm: auto
        fixed_position: true
        allow_fixed_grasp: true
        force_fixed_grasp: true
    mode: manual
    planner: curobo
  instruction: put the banana on the top of the vintage rusty frying pan.
  layout_config: &id003
    ignored_objects: []
    type: None
  mode: manual
  num_episode: 1
  object_config: &id004
    '0':
      type: existed_object
      uid_list:
      - fb1b6fc41f7e49adbf467e5e5988d190
    '1':
      type: existed_object
      uid_list:
      - 1fdc84a7be2c4348b281490c89d76062
  preprocess_config:
  - &id005
    config:
      config: default
      type: convexDecomposition
    type: collider
  - type: ccd
  robots:
  - &id006
    config:
      gripper_type: aloha_split
    type:  aloha_split
  table_uid: aa49db8a801d402dac6cf1579536502c
  task_name: Minimal_Banana
  usd_name: scene_usds/debug_scenes/banana_plate_scenes/base
evaluation_configs:
- domain_randomization: *id001
  generation_config: *id002
  instruction: put the banana on the top of the vintage rusty frying pan.
  layout_config: *id003
  mode: manual
  num_test: 10
  object_config: *id004
  preprocess_config:
  - *id005
  robots:
  - *id006
  table_uid: aa49db8a801d402dac6cf1579536502c
  task_name: Minimal_Banana
  usd_name: scene_usds/debug_scenes/banana_plate_scenes/base

[Axi404]

This issue is caused by the action space constraint. In demogen, a normal action consists of 6 phases:

• 0, 1, 2: pre_grasp, grasp, post_grasp
• 3, 4, 5: pre_place, place, post_place

In most cases, motion planning failures are caused by the target pose being outside the valid action space.
In this specific case, the place pose is out of the action space.
In addition, the dual-arm setup has a more limited action space than the single-arm setup, so there is currently no automatic hand-over skill.
As a result, you need to adjust the layout manually.
I adjusted the layout configuration (you can also directly edit the layout in the USD file), try it:

demonstration_configs:
- domain_randomization: &id001
    cameras:
      config_path: configs/cameras/fixed_camera_aloha_refined.yml
      type: fixed
    random_environment:
      has_wall: false
      hdr: false
      robot_base_position: false
      robot_eepose: false
      table_texture: false
      table_type: false
      wall_texture: false
    rewrite_instruction: false
  generation_config: &id002
    action_path:
      mode: auto
      robot: 0
    articulation: []
    goal:
    - - obj1_uid:
        - '0'
        obj2_uid:
        - '1'
        position:
        - top
        arm: auto
        fixed_position: true
        allow_fixed_grasp: true
        force_fixed_grasp: true
    mode: manual
    planner: curobo
  instruction: put the banana on the top of the vintage rusty frying pan.
  layout_config: &id003
    ignored_objects: []
    custom_tableset:
      '0':
        pos_x: 0.08831
        pos_y: 0.0631
        random_range_angle: 0
        reset_orientation: true
        orientation:
        - 0.0
        - 0.0
        - 1.0
        - 0.0
        type: fixed_global_range
      '1':
        pos_x: 0.08
        pos_y: 0.34616
        random_range_angle: 0
        reset_orientation: true
        type: fixed_global_range
    type: random_custom_tableset
  mode: manual
  num_episode: 100
  object_config: &id004
    '0':
      type: existed_object
      uid_list:
      - fb1b6fc41f7e49adbf467e5e5988d190
    '1':
      type: existed_object
      uid_list:
      - 1fdc84a7be2c4348b281490c89d76062
  preprocess_config:
  - &id005
    config:
      config: default
      type: convexDecomposition
    type: collider
  - type: ccd
  robots:
  - &id006
    config:
      gripper_type: aloha_split
    type:  aloha_split
  table_uid: aa49db8a801d402dac6cf1579536502c
  task_name: Minimal_Banana
  usd_name: scene_usds/debug_scenes/banana_plate_scenes/base
evaluation_configs:
- domain_randomization: *id001
  generation_config: *id002
  instruction: put the banana on the top of the vintage rusty frying pan.
  layout_config: *id003
  mode: manual
  num_test: 10
  object_config: *id004
  preprocess_config:
  - *id005
  robots:
  - *id006
  table_uid: aa49db8a801d402dac6cf1579536502c
  task_name: Minimal_Banana
  usd_name: scene_usds/debug_scenes/banana_plate_scenes/base

[jbpen]

You mentioned this point “there is currently no automatic hand-over skill”，does this mean that GenManip currently does not support dual-arm collaboration?

[Axi404]

It depends on your definition. For example, in the existing best practices, GenManip can use two arms to sequentially perform different operations to complete tasks such as making a sandwich. In addition, you can implement your own skill to support more advanced behaviors.