Digital twin of a 7-DoF robotic arm

This is a contribution to the NBEL lab at the Open University of Israel, mainly a dataset-generation tool that other students later used in their research. It is a digital twin of a 7-DoF robotic arm that couples MuJoCo for physics with Unity for rendering, so the same motion can be simulated, visualized, and recorded as training data.

The practical goal is a controllable dataset generator for dynamic scenes. Real robot time is scarce, and event-based cameras make that worse: you need lighting, motion speed, object appearance, and polarity thresholds to vary together. Here the user sets those knobs in a Unity GUI — object type and color, scale, background, arm speed, event-camera positive and negative thresholds, lighting intensity and angle, and recording length — or a randomizer draws them uniformly for a batch of scenes. Each take is stored as a pickle whose filename encodes the chosen settings, so a learning pipeline can recover the ground-truth configuration without a separate metadata file.

That split of engines matters. MuJoCo gives a physically consistent arm (forward and inverse kinematics live in the Python stack, with a path toward the real VX300s / Dynamixel hardware). Unity supplies the camera, materials, and lighting that a perception model actually sees. gRPC ties the two, which is what a digital twin has to do: one robot state, two views of the world.

The importance is not the demo arm by itself. Labeled, randomized, physics-backed video of a manipulator in clutter is expensive to collect on hardware and hard to reproduce. A twin that can also emulate event-camera thresholds lets you study perception and control before you burn lab time, and it gives a shared artifact — code, recordings, and a public video — for anyone who wants to extend the NBEL setup.

Code: https://github.com/Turgibot/Digital_Twin_7DoF_Robotic_Arm

Video: https://www.youtube.com/watch?v=y0JEVJ6KaMs