Tournaments (Optimization)

General

Tournaments are rigorous, focused competitions centered on "Benchmarking & Policy Improvement."

• Purpose: To drive optimization and identify the highest-performing robotic control policies.
• Scope: Specific environments with defined parameters and specific robotic agents(e.g., “Nova Carter robot in a warehouse moving from random position A to B”).
• Metrics: Agents are scored based on objective measurements, including speed, accuracy, and overall success Rate.
• Validation: Validation of each submitted agent is transparently benchmarked using blockchain technology.
• Reward: Reward distribution is a winner-takes-all approach. However, the platform will payout bounties to extra community members who have made meaningful contributions to the tournament(e.g., help platform to improve evaluation approach, bug-fix and very slightly lower than the winner).
• Open-sourced: Tournaments result will be published from its code to checkpoints for accelerating Robotics AI.

Compete

• RL algorithm (Reward function, reward weights)
• Training methodology
• Train dataset

Participation

Tournaments will be organized as soon as a new tournament task is ready. We aim to organize tournaments sequentially, enabling robotics engineers to participate in tournaments based on their availability and experience level.

When a new tournament is ready, the relevant GitHub repository will be shared, containing the 3D environment and robot assets in USD and URDF formats, along with public evaluation logic.

Generally, we aim to design tournament submissions as Isaac Lab External Projects, which is the standard approach for developing robot policies in Isaac Lab. As tasks differ between tournaments, we will release task-specific documentation for each GitHub repository.

Breakdown

One single tournament is consisted of below sessions.

• Contest session:: A session that miners optimize their robot policy for given task. Since tournament task is created on tournament platform until submit session.
• Submit session:: A session that allows miners submit their project.
• Evaluation session:: A session that validators evaluate submissions that were submitted during submit session.
• Review session:: A session that top submission is being reviewed by owner team to identify cheating submission and select the honest top submission.
• Reward session:: A session that top miner get rewards in subnet alpha token.