Robotics Simulation, Digital Twins & Training
Unitree G1-boxer
A Unitree G1 humanoid trained from real boxing footage to reproduce footwork, weight transfer and full-body shadowboxing in NVIDIA Isaac Lab — then brought into Isaac Sim for cinematic presentation.
The problem
Human boxing is more than an arm animation. A convincing straight, step, weight transfer or defensive movement depends on balance, foot placement, body rotation and timing working together as one physical motion. The challenge was to take real shadowboxing footage and teach a Unitree G1 humanoid to reproduce that movement through physics.
What we solved
Real boxing footage was reconstructed and retargeted into a G1-compatible motion reference, then used to train a whole-body control policy in NVIDIA Isaac Lab. The workflow was developed progressively: first teaching clear forward, backward and lateral footwork, then cleaner punch mechanics and finally longer shadowboxing sequences combining movement, balance and striking. The training setup was refined to preserve both motion fidelity and stability, with particular focus on spatial movement, weight transfer, punch trajectory and full-body coordination. The trained policy was then transferred into NVIDIA Isaac Sim, baked into a production-ready animation and integrated into a fully lit boxing-gym environment for final rendering.
The result
A physics-driven Unitree G1 boxer capable of completing full shadowboxing sequences while maintaining balance, deliberately moving through space and reproducing recognisable boxing mechanics including punches, moving the head off the center line, lateral movement, weight transfer and recovery. The final motion is not a traditional keyframed character animation. It originates from a learned control policy operating a simulated humanoid, then becomes part of a cinematic 3D production pipeline for presentation.