Skip to content

phases.md

Latest commit

 

History

History
68 lines (44 loc) · 3.68 KB

phases.md

File metadata and controls

68 lines (44 loc) · 3.68 KB

Competition Phases

Qualification Phase: Train Your Model

During qualification, participants use their preferred tools—including open source software and simulators—alongside the Intrinsic challenge toolkit to train a model for the cable insertion task. All submitted models are evaluated using Gazebo.

Technical Overview

Review the core technical requirements for this phase, including setup constraints, trial descriptions, and evaluation metrics. Refer to the Qualification Phase: Technical Overview document for full specifications.

Implementation Workflow

To successfully qualify, participants must create a ROS 2 node that adheres to the behavioral requirements specified in the Challenge Rules.

Recommended Approach: Using the aic_model Framework

For convenience, we provide an aic_model framework that handles all ROS 2 lifecycle management and boilerplate. You simply implement a Python policy class:

  1. Create a Policy Class: Define a Python class that derives from Policy.
  2. Implement insert_cable(): This method is called when aic_engine requests a new task. It receives observation data and callable methods for robot control.
  3. Load Your Model: Initialize your trained policy (e.g., PyTorch checkpoint, ONNX model, or control algorithm) when your class is instantiated.
  4. Process Observations: Use the provided get_observation() callback to retrieve sensor data at up to 20 Hz.
  5. Output Commands: Use move_robot() and other provided methods to command the robot.
  6. Return on Completion: Your insert_cable() method should return when the task is complete.

Tutorial: For a step-by-step guide, see Creating a New Policy Node.

Example: Reference implementation: WaveArm.py

Alternative: Implement Your Own Node

You may also implement your own ROS 2 node from scratch, as long as it:

  • Is named aic_model and implements the ROS 2 Lifecycle interface
  • Responds to the /insert_cable action server
  • Follows all requirements in the Challenge Rules

Participation Guidelines

  • Policy Development: Participants are free to use any approach to develop a policy, including:
    • Real-world teleoperation data.
    • Training in a simulator of choice (MuJoCo, Isaac Sim, O3DE, etc.).
    • Classical control algorithms.
  • Interface Requirements: Policies (wrapped in the service described above) must consume world information and output actions using standard formats.
  • Evaluation: The provided Evaluator Simulator (Gazebo) scores the performance of participant models.
    • During development, participants can run the Evaluator Simulator locally to test performance.
    • Upon submission, a cloud instance runs the same Evaluator Simulator to log official scores.

For more information, please refer to:

Phase 1: Develop in Flowstate

Teams advancing to Phase 1 gain access to Intrinsic Flowstate (our development environment) and the Intrinsic Vision Model. Using these tools, teams will build a complete robotic cable handling solution incorporating their trained models.

Coming Soon

Phase 2: Run on Real Robots

Phase 2 participants deploy their solutions to a physical robotic workcell at Intrinsic’s HQ. This phase validates solutions in the real world and determines prize winners.

Coming Soon