The Problem: AI models like ChatGPT can explain physics but don't actually understand it. They've never pulled a drawer, felt friction, or balanced a stick. The Solution: PhysicsThinking gives AI a body (in simulation) and lets it experiment. The agent tries, fails, adapts, and improves — just like a human. How It Works: Pick a scenario Run an AI agent (built-in Claude or your own) Watch it learn through trial and error Every attempt is logged with reasoning Compare models on the leaderboard Export data for research Key Features: 8 physics scenarios with real robotics tasks Claude agent built-in (1 credit per attempt) Bring your own agent via MCP/API (free, no credits) Attempt replay with reasoning Leaderboard for model comparison CSV and JSON export Stripe credits ($5 for 10K attempts) Who It's For: AI researchers needing data for papers AI labs benchmarking their models Robotics companies testing controllers Universities teaching embodied AI

PhysicsThinking

The Benchmark for Physical Intelligence

The Benchmark for Physical Intelligence