The AI for Industry Challenge Toolkit is LIVE!
Click here to access the Github reposiutory.
What’s inside the box?
We are handing you a high-fidelity environment tuned to match real-world conditions:
The scene: A complete robotic workcell including the enclosure, lighting, and room assets. The manipulator is a Universal Robots UR5e equipped with an ATI Industrial Automation force-torque sensor, 3 wrist-mounted Basler AG cameras, and a Robotiq Hand-E gripper.
The challenge assets: We’ve included high-quality assets for electronics assembly, including network cards, flexible cables and our task board.
The backbone: We’re leveraging ROS for seamless data flow and hardware abstraction, paired with Zenoh (rmw_zenoh) for performant data transport.
High-performance robot control: We’ve developed and released a specialized Impedance controller that responds to both cartesian and joint configuration inputs. It handles precise control at 500Hz for smooth motions while remaining capable of receiving targets at 10-30Hz frequencies (e.g., from a VLA policy).
Multi-simulator support: By default, the scene loads in Gazebo, which serves as our official evaluation environment. However, we also provide the flexibility to launch the same scene in MuJoCo (via our partnership with Google Deepmind) or as a USD in Isaac Lab for massive parallel training.
Cross-platform support: We’re relying on pixi to ensure all participant-facing packages can be installed seamlessly on any system.
Data collection & teleop: We’ve integrated with the popular Lerobot framework from Hugging Face, making it possible to immediately reuse world-class tooling to teleoperate, collect data, and train your policies.
If you have any questions, please start a new discussion thread in AI for Industry Challenge - Open Robotics Discourse.
We’re really excited for everyone’s participation!