ROBOTIS Teaches Humanoid Robot K-POP Dance from Video

Reporting by Interesting Engineering describes that ROBOTIS demonstrated its open-source AI Sapiens humanoid platform learning the CORTIS REDRED Challenge dance using only smartphone video rather than professional motion-capture systems. Reporting by Interesting Engineering and Mike Kalil says the demonstration combined video-based motion capture, motion retargeting, simulation-based reinforcement learning, and Sim2Real transfer to move learned behaviors from a simulated digital twin to a physical robot. Mike Kalil reports the robot measures 1.3 meters in height, weighs 34 kg, has 23 degrees of freedom, and runs onboard inference on an 8-core ARM CPU with a Mali GPU and a dedicated NPU; Kalil also notes ROBOTIS has not confirmed a release date or pricing, though Korean reports cited by Kalil place potential price below $10,000.

The public descriptions place the pipeline stages that matter to practitioners into four parts: smartphone video capture; pose extraction and motion retargeting to a simulated humanoid; reinforcement learning in simulation to optimize balance and timing; and Sim2Real transfer for deployment on hardware. Reporting highlights the use of DYNAMIXEL-Q actuators in the physical robot, which are presented as part of the hardware stack supporting the transfer of learned policies to actuated joints.

For robotics researchers and engineers, the two notable trends in this demonstration are the use of commodity capture (smartphone video) to generate training targets and the publication of an open-source physical-AI framework. Comparable public demonstrations increasingly emphasize lowering setup cost for imitation learning; that pattern accelerates experimentation because teams can iterate on motion datasets without access to studio-grade mocap. The reported hardware specs make AI Sapiens roughly comparable on paper to other recent lower-cost humanoids, though public reporting does not provide benchmarking data on stability, repeatability, or sample efficiency.

Observers should look for a public code or dataset release, technical documentation of the retargeting and RL reward design, and controlled evaluations comparing sim-to-real robustness to existing platforms. Pricing and availability statements from ROBOTIS would clarify whether this demonstration translates into a broadly accessible research platform. Additionally, independent tests measuring repeatability across runs and robustness to variations in phone-camera quality will be key indicators of practical utility.