Humanoids Copy Pro Athletes' Style: Planner and Tracker Split, Tested on 2 Robots

The team behind the Noitom Robotics white paper proposes a way to teach a humanoid robot to move like a professional athlete without losing accuracy on the task. The method was tested on two humanoid platforms. One of them is full-size, and it performed a serve "in the wild," without a motion capture system. This is the only test mentioned in the announcement. The full text sits behind a download form on the Wiley site, and the public description contains no figures.

What's the problem

In ball-sport robotics, style and outcome pull against each other. If you copy an athlete's movements as closely as possible, the robot often misses the task: its masses, moments, actuator delays and joint limits are different. If you optimize only for hitting the target, the motion works but does not look human. The authors call this an open problem in the field.

Their answer is an architectural split. A high-level planner decides which motion to perform and when. A low-level tracker makes sure the robot's body physically reproduces that motion and does not fall over.

Where the references come from

The data sources are named outright: television broadcasts and motion capture recordings. Pose and trajectory are reconstructed from video, then retargeted to the humanoid's kinematics. Physical correction follows, which removes what a human can do and a robot cannot: unrealistic joint speeds, loss of contact with the support surface, loss of balance. The output is reference motions that the robot can track safely.

Using broadcasts noticeably lowers the cost of data collection. There is no need to bring an athlete into a studio with dozens of cameras; archived matches are enough. The quality of reconstruction from a single camera remains an open question, and the announcement does not answer it.

Speed as a control parameter

The second part of the work concerns the transition from simulation to reality. Tracking errors accumulate there: a small lag in one step shifts the next state, and by the moment of impact the discrepancy can become critical. The authors propose two steps:

• train the tracker on random execution speeds of the reference motion;
• add a trained speed adapter to the planner, which sets the tempo according to the robot's current state.

Essentially, time becomes a controllable variable. If the robot cannot keep up with the reference, the planner slows down or speeds up the reference rather than forcing the tracker to catch up at any cost.

What is known about the test

• Platforms: two humanoids.
• One of them is full-size.
• Task: a serve performed outside the lab.
• No motion capture system was used in the test.
• Quantitative metrics (hit accuracy, fall rate, tracking error) are not disclosed in the public part.

How to read this material

This is a marketing page for a white paper, not a peer-reviewed article, and you cannot judge the result from it. What's needed are concrete numbers: how much the speed adapter reduces error compared with a tracker trained at a fixed tempo, and how many attempts out of a hundred end in a successful serve.

If you are building control for humanoids, the most useful takeaway is the idea of decoupling. You can test it yourself without anyone else's hardware: run a tracker at different reference speeds in simulation and see how the error grows. If the growth is steep, speed randomization during training will give a quick win. Before building the method into a development plan, download the document and check its claims against the results tables.