RoboColosseum

The proving ground for physical AI.

🤗 Explore rollout data   ·   🏆 Leaderboard   ·   🌐 Website   ·   📄 Paper   ·   💻 GitHub

RoboColosseum evaluates robot policies in the real world, at scale. We measure how reliably models work across everyday environments, tasks, and robot embodiments.

• 🤖 100+ robots, including SO-101, Franka, bimanual YAM, and Unitree G1
• 🌍 140+ real-world scenes with thousands of tasks
• 🎓 10,000 free evaluation hours in a shared pool for academic labs

🤗 Open rollout data

Our commitment: every evaluation rollout, openly released on Hugging Face—failures included.

Explore robot attempts in the datasets below to understand policy behavior and failures. Each dataset card explains its contents, evaluation setup, and usage details.

🛤️ Two evaluation tracks

Open track. Blind, head-to-head tests measure how well models handle new tasks in real-world environments, with no task-specific training.

Fine-tune track. Models learn from the same fixed, open-source dataset, then undergo evaluation to measure how well they adapt to the task.

🎓 Free evaluations for academia

Need evaluations beyond what your lab can run? Apply for a share of 10,000 free evaluation hours →