RoboChallenge-CVPR-2026

reproduce steps

1. download model, python download_checkpoints.py
2. edit submission ids in Makefile
3. eval a robot, make eval_pi05_{robot_type}

---

CVPR Pi0.5 Multi-Task Example

This branch is a CVPR on-robot inference example for Pi0.5 multi-task policies. There is one checkpoint per robot platform; you can evaluate all tasks supported on that platform with the same model, by changing the task prompt (and episode data) only. Supported robot types in this codebase (see demo.py / test.py): dosw, aloha, arx5, ur5.

Core entrypoints:

• test.py — Connects to the local mock robot server for quick debugging.
• demo.py — RoboChallenge official evaluation (requires user_token and submission_id).

Quick start

1. Installation

uv venv
source .venv/bin/activate   # if your uv venv uses another path, activate that instead
uv pip install -e ./openpi
uv pip install pytest
uv pip install -r requirements.txt

2. Download the example model

Download the Pi0.5 multi-task checkpoint from RoboChallenge on Hugging Face. One checkpoint per embodiment covers all tasks on that robot; switch tasks via --prompt only.

3. Run the mock test

1. Edit mock_server/mock_settings.py: set ROBOT_TAG and RECORD_DATA_DIR (only one active pair). With the bundled sample data (paths relative to repo root):

   • ROBOT_TAG='aloha', RECORD_DATA_DIR='../20260413/aloha/pack_the_toothbrush_holder'
   • ROBOT_TAG='w1', RECORD_DATA_DIR='../20260413/w1/sweep_the_trash' → run test.py with --robot_type dosw
   • ROBOT_TAG='ur5', RECORD_DATA_DIR='../20260413/ur5/arrange_fruits'
   • ROBOT_TAG='arx5', RECORD_DATA_DIR='../20260413/arx5/hang_the_cup'

2. Terminal 1 — start the mock server:

   cd mock_server
   python3 mock_robot_server.py
   

   For debug mode (state includes a consecutive action_chunk starting from current frame, up to 30 frames total), start with:

   cd mock_server
   python3 mock_robot_server.py --debug
   

3. Terminal 2 — from the repo root:

   python3 test.py \
     --checkpoint /path/to/checkpoint_dir \
     --prompt "task instruction in natural language(match the format used in training)" \
     --robot_type dosw
   

   Use exactly one of --robot_type dosw | aloha | arx5 | ur5 so it matches ROBOT_TAG.

   Adjust --action_type, --duration, --image_size, etc. as needed (see test.py).

4. Run the demo

After you have a submission and are in the assigned evaluation window, run demo.py.

W1

python3 demo.py \
  --user_token <your_user_token> \
  --submission_id <your_submission_id> \
  --checkpoint /path/to/w1_checkpoint_dir \
  --prompt "task instruction in natural language(match the format used in training)" \
  --action_type joint \
  --image_size "640x480" \
  --robot_type dosw

Aloha

python3 demo.py \
  --user_token <your_user_token> \
  --submission_id <your_submission_id> \
  --checkpoint /path/to/aloha_checkpoint_dir \
  --prompt "task instruction in natural language(match the format used in training)" \
  --action_type joint \
  --image_size "640x480" \
  --robot_type aloha

ARX5

python3 demo.py \
  --user_token <your_user_token> \
  --submission_id <your_submission_id> \
  --checkpoint /path/to/arx5_checkpoint_dir \
  --prompt "task instruction in natural language(match the format used in training)" \
  --action_type leftjoint \
  --image_size "1280x720" \
  --robot_type arx5

UR5

python3 demo.py \
  --user_token <your_user_token> \
  --submission_id <your_submission_id> \
  --checkpoint /path/to/ur5_checkpoint_dir \
  --prompt "task instruction in natural language(match the format used in training)" \
  --action_type leftjoint \
  --image_size "640x480" \
  --robot_type ur5

---

CVPR DM0 Multi-Task Example

This section covers using the DM0 policy backend. The entry points are:

• test.py — Mock robot server for local debugging (--policy_type dm0).
• eval.py — RoboChallenge official evaluation.

Quick start (DM0)

1. Installation

Same as Pi0.5. See the Installation section above.

2. Run the mock test (DM0)

1. Configure mock_server/mock_settings.py the same way as the Pi0.5 section above.

2. Terminal 1 — start the mock server:

   cd mock_server
   python3 mock_robot_server.py
   

3. Terminal 2 — from the repo root:

   python3 test.py \
     --policy_type dm0 \
     --checkpoint /path/to/dm0_checkpoint_dir \
     --prompt "task instruction in natural language" \
     --robot_type arx5
   

   Add --save_state to save every state observation to eval_data/:

   python3 test.py \
     --policy_type dm0 \
     --checkpoint /path/to/dm0_checkpoint_dir \
     --prompt "task instruction in natural language" \
     --robot_type arx5 \
     --save_state
   

3. Run the official evaluation (DM0)

Use eval.py instead of demo.py for the DM0 backend.

W1

python3 eval.py \
  --user_token <your_user_token> \
  --submission_id <your_submission_id> \
  --checkpoint /path/to/dm0_w1_checkpoint_dir \
  --prompt "task instruction in natural language" \
  --action_type joint \
  --image_size "640x480" \
  --robot_type dosw

Aloha

python3 eval.py \
  --user_token <your_user_token> \
  --submission_id <your_submission_id> \
  --checkpoint /path/to/dm0_aloha_checkpoint_dir \
  --prompt "task instruction in natural language" \
  --action_type joint \
  --image_size "640x480" \
  --robot_type aloha

ARX5

python3 eval.py \
  --user_token <your_user_token> \
  --submission_id <your_submission_id> \
  --checkpoint /path/to/dm0_arx5_checkpoint_dir \
  --prompt "task instruction in natural language" \
  --action_type leftjoint \
  --image_size "1280x720" \
  --robot_type arx5

UR5

python3 eval.py \
  --user_token <your_user_token> \
  --submission_id <your_submission_id> \
  --checkpoint /path/to/dm0_ur5_checkpoint_dir \
  --prompt "task instruction in natural language" \
  --action_type leftjoint \
  --image_size "640x480" \
  --robot_type ur5

Add --save_state to any of the above commands to save:

• state observations: eval_data/{robot_id}/{date}-{time}/{job_id}/state_{seq}.pkl
• posted actions: eval_data/{robot_id}/{date}-{time}/{job_id}/infer_{seq}.json

---

Original README.md

RoboChallengeInference

Project Structure

- RoboChallengeInference/
    - README.md
    - requirements.txt
    - demo.py
    - test.py  # Main test entry script
    - robot/
        - __init__.py
        - interface_client.py
        - job_worker.py
    - mock_server
        - mock_rc_robot.py
        - mock_robot_server.py
        - mock_settings.py
        - utils.py
    - utils/
        - __init__.py
        - enums.py
        - log.py
        - util.py

User Guide

1. Installation

# Clone the repository and checkout the specified branch
git clone https://github.com/RoboChallenge/RoboChallengeInference.git
cd RoboChallengeInference

# (Recommended) Create and activate a virtual environment to avoid polluting your global Python environment
python -m venv venv
source venv/bin/activate

# Install dependencies
pip install -r requirements.txt

2. Checkout & Modification

# Checkout
git checkout -b my-feature-branch
# Follow the instructions in demo.py to modify parameters and implement your custom inference logic based on DummyPolicy.
# The current task prompt will be passed into `DummyPolicy.run_policy(input_data, prompt=...)`.

3. Test

# Open the mock_settings.py file and set the ROBOT_TAG and RECORD_DATA_DIR variables according to your robot and data directory requirements.
# Notes:
#   Only one pair of ROBOT_TAG and RECORD_DATA_DIR should be active at a time.
#   Ensure that the RECORD_DATA_DIR path matches the structure of your data folder.
#   You can find the appropriate ROBOT_TAG in your training data or on our website.
#   For the 20260413 CVPR package, you can use one of the following pairs:
#     ROBOT_TAG='aloha', RECORD_DATA_DIR='../20260413/aloha/pack_the_toothbrush_holder'
#     ROBOT_TAG='w1',    RECORD_DATA_DIR='../20260413/w1/sweep_the_trash'
#     ROBOT_TAG='ur5',   RECORD_DATA_DIR='../20260413/ur5/arrange_fruits'
#     ROBOT_TAG='arx5',  RECORD_DATA_DIR='../20260413/arx5/hang_the_cup'
#   RECORD_DATA_DIR also supports robot-level directory (e.g. '../20260413/ur5').
#   The mock server will auto-detect the task directory with meta/states/videos.
# Start the test service
cd mock_server
python3 mock_robot_server.py
# Use test.py for testing; it will automatically invoke the mock interface to help you debug your model
# Replace {your_args} with the actual parameters you want to test, for example: --checkpoint xxx.
# Run this in another shell at repo root.
python3 test.py {your_args}

4. Submit

• Log in to RoboChallenge Web
• Submit an evaluation request
• On the "My Submission" page, you can view your submissions. Click "Detail" to see more information about a submission.
• The Submission ID displayed on the details page will be required for the evaluation process. The program will automatically poll and select active runs under that submission.

5. Execute

• Wait for a notification (on the website or via email) indicating that your task has been assigned.
• Ensure the modified code from the previous steps is actively running during the assigned period.
• Start the evaluation worker with:

python3 demo.py --user_token <your_user_token> --submission_id <your_submission_id> --checkpoint <your_checkpoint>

• After the task is completed, the program will exit normally. If you encounter any issues or exceptions, please feel free to contact us.

6. Result

Once your task has been executed, you can view the results by visiting the "My Submissions" page on the website.

Key API Parameter Descriptions

This is the direct interface for the robot. The base URL is /api/robot/<id>/direct. For example, if the robot ID is 1, the full URL to get the state is /api/robot/1/direct/state.pkl.

Sync Clock

Endpoint: /clock-sync
Method: GET

Request Parameters

None

Response Example

{
  "timestamp": 0.0
}

Response Fields

---

Get State

Endpoint: /state.pkl
Method: GET

Request Parameters

Robot-specific image_type values and returned images keys:

Robot-specific action_type values:

Response Example

The response is a pickle file containing a dictionary with the following structure. The images keys depend on robot type:

• aloha / w1

{
    "images": {
        "cam_left_wrist": b'PNG',
        "cam_right_wrist": b'PNG',
        "cam_high": b'PNG'
    }
}

• ur5

{
    "images": {
        "cam_global": b'PNG',
        "cam_arm": b'PNG'
    }
}

• arx (arx5)

{
    "images": {
        "cam_global": b'PNG',
        "cam_arm": b'PNG',
        "cam_side": b'PNG'
    }
}

Full response example (aloha, action_type=joint):

{
    "state": 'normal',
    "timestamp": 1774949968.382069,
    "pending_actions": 0,
    "action": [
        -0.0151, 0.0, 0.0, -0.0184, 0.0986, 0.0529, 0.0001,
         0.0113, 0.0, 0.0, -0.0079, 0.0962, 0.0298, 0.0
    ],
    "images": {
        "cam_left_wrist": b'PNG',
        "cam_right_wrist": b'PNG',
        "cam_high": b'PNG'
    }
}

Response Fields

---

Post Action

Endpoint: /action
Method: POST

Request Parameters

The HTTP body should be a JSON object with the following structure:

{
  "actions": [
    [
      0.0,
      0.0
    ],
    [
      0.0,
      0.0
    ],
    [
      0.0,
      0.0
    ]
  ],
  "duration": 0.0
}

Response Example

{
  "result": "success",
  "message": ""
}

Response Fields

Robot specific Notes

Different robots have different action shapes and camera placement.

• W1

  • Dual-arm robot
  • 7 DOF per arm (6 joints + 1 gripper)
  • Joint control:
    • one arm(left or right): 7 numbers total: [6 joints, 1 gripper]
    • two arms: 14 numbers total: [left 6 joints, left 1 gripper, right 6 joints, right 1 gripper]
  • Pose control
    • one arm(left or right): 8 numbers total: [x, y, z, quaternion(xyzw), gripper]
    • two arms: 16 numbers total: [left x, left y, left z, left quaternion(xyzw), left gripper, right x, right y, right z, right quaternion(xyzw), right gripper]
  • 3 cameras: mounted on left/right arm, and on the top of the robot

• Aloha

  • Dual-arm robot
    • 7 DOF per arm (6 joints + 1 gripper)
    • Joint control:
      • one arm(left or right): 7 numbers total: [6 joints, 1 gripper]
      • two arms: 14 numbers total: [left 6 joints, left 1 gripper, right 6 joints, right 1 gripper]
    • Pose control
      • one arm(left or right): 8 numbers total: [x, y, z, quaternion(xyzw), gripper]
      • two arms: 16 numbers total: [left x, left y, left z, left quaternion(xyzw), left gripper, right x, right y, right z, right quaternion(xyzw), right gripper]
    • 3 cameras: mounted on left/right arm, and on the top of the robot

• Arx5

  • Single-arm robot
  • 7 DOF (6 joints + 1 gripper)
  • Joint control: 7 numbers total: [6 joints, 1 gripper]
  • Pose control: 8 numbers total: [x, y, z, quaternion(xyzw), gripper]
  • You must always use left in the action_type parameter, e.g., leftjoint or leftpos.
  • 3 cameras: mounted on the arm, opposite to the arm, and on the right side of the arm

• Ur5

  • Single-arm robot
  • 7 DOF (6 joints + 1 gripper)
  • Joint control: 7 numbers total: [6 joints, 1 gripper]
  • Pose control: 8 numbers total: [x, y, z, quaternion(xyzw), gripper]
  • You must always use left in the action_type parameter, e.g., leftjoint or leftpos.
  • 2 cameras: mounted on the arm, and opposite to the arm

Contact

For official inquiries or support, you can reach us via:

• GitHub Issues: https://github.com/RoboChallenge/RoboChallengeInference/issues
• Reddit: https://www.reddit.com/r/RoboChallenge/
• Discord: https://discord.gg/8pD8QWDv
• X (Twitter): https://x.com/RoboChallengeAI
• HuggingFace: https://huggingface.co/RoboChallenge
• GitHub: https://github.com/RoboChallenge
• Support Email: [email protected]