PhysicsLens

Every robot arm is slightly different from its simulator: a little delay, some friction, a deadband in the servos, damping nobody measured. A policy trained in simulation then fails on the real arm. Fixing it normally takes a lab rig and an expert. PhysicsLens does it from a video or from the robot's own recordings. For robots: - Robot calibration: give it an arm's recordings, and it identifies a servo law for every joint (delay, friction, deadband, damping), writes them into the arm's MuJoCo model, and checks the corrected simulator against episodes it never saw. Works on the SO-100 and the ALOHA arms, from the web app or the command line. - Sim vs real: the real arm and its corrected simulator side by side, with each model's error against the real joints over time. - Per-arm and over time: calibrate each unit separately, and compare recordings months apart to see when an arm needs recalibrating. - Compensator: the laws, inverted, correct the commands sent to the arm. For anything that moves: Film it or upload a clip, and an AI agent finds the differential equation that governs the motion. - Scene understanding: Muse Spark Open looks at a few keyframes, names the moving object and the kind of motion, and finds a reference object for scale. - Tracking: SAM 3.1 follows the object through the video and turns its masks into position, orientation, and real units. - Hypotheses: the agent proposes candidate laws, and writes its own as an equation when no textbook law fits. - Verification: code fits every candidate and accepts only the law that forecasts held-out motion best. The language model never declares an equation correct. - Experiment requests: when two laws explain the data equally well, the agent designs the experiment that would separate them and asks you for it, with Muse Image drawing the setup on your own frame. If nothing fits, it says so instead of forcing an answer. - Live mode: the equation converges on screen while you film, and it flags when the dynamics change. What you get: the equation with its parameters and uncertainty, a simulated "ghost" of the law playing over your video, a decision trail showing what the agent saw, proposed, rejected, and asked for (with the model behind every step), voice what-ifs while filming ("double the damping") through Muse Voice Transcribe, and exports: MuJoCo model, standalone Python simulation, JSON, and a shareable video. Also in the app: - Solar System Lab: a 3D solar system on our own N-body engine; change it (make Jupiter heavier), and the agent rediscovers the laws of the universe you made. - Physics audit: checks whether a video, real or AI-generated, obeys physics, and gives the measurement behind every verdict. - Flight: thrust and drag laws for drones and rockets from logs and launch video. - MCP server, so other AI assistants can call PhysicsLens as a tool. Every reasoning step runs on open weights: Muse Spark Open on Baseten, with Muse Glimmer self-hosted as the second path.