Teaching humanoid robots to learn movement by watching humans.
Movement Lab is a build-in-public project working toward a humanoid that can observe human motion — dance, athletics, everyday dexterity — and learn to reproduce it. The long-term north star is a robot that picks up movement the way a person does: by watching, mimicking, and refining.
This repo is stone one of that vision — scoped small, built honestly, one joint at a time.
A humanoid that learns physical skill from human demonstration: watch a salsa lead, a jump, a gesture — retarget it onto the robot's body, refine it in simulation, and eventually run it on real hardware. Ambitious on purpose. Built incrementally on purpose.
Early foundation, and growing in the open:
- Motion fundamentals in NumPy — poses, trajectories, velocity, acceleration
- First robot built, rendered, and animated in MuJoCo (a single-joint arm, swinging under generated motion)
- Next: multi-joint arms, then retargeting human motion onto a humanoid model
The pipeline this is building toward, the same one behind modern humanoid skill learning:
human demonstration → retarget onto robot → refine in simulation → real hardware
Core working principle: generate → simulate → judge. Code generates motion, simulation surfaces problems cheaply before any hardware contact, and human judgment decides whether the result is actually right.
Python · NumPy · MuJoCo · Google Colab
Built by a professional humanoid-robot teleoperator — the demonstration data this whole field learns from is what I work with daily. Movement Lab is where that hands-on motion intuition meets the technical stack.