Humanoid general motion tracking — getting a robot to follow arbitrary whole-body reference motions while keeping its balance — involves dozens of practical design choices whose individual effects are rarely isolated. A new paper from IIT's HUCEBOT team (Fabio Amadio and Enrico Mingo Hoffman), submitted to arXiv on July 22, does the unglamorous but essential work: a controlled, reproducible empirical study of what actually moves the needle, built around YAHMP, a new open-source modular framework for training, evaluating and deploying whole-body tracking policies on the Unitree G1.
Six Knobs, Systematically Turned
YAHMP defines a nominal configuration and then varies one factor at a time: motion-command representation, observation history, action representation, actuation profile, hand-force randomization during training, and training approach. Policies are evaluated on a test set of retargeted human motions, with TWIST2 as an external baseline trained on the same motion set — a rare apples-to-apples comparison in a field where most papers report only self-baselines.
The results cleanly separate two families of choices. Some — like command and action representations — have clear tracking effects. Others mainly change actuation effort, training complexity, or physical interaction capability without meaningfully improving tracking accuracy. Compared with the nominal YAHMP policy, TWIST2 achieves 10% lower key-body position error but pays for it elsewhere: base errors 32-36% higher, key-body orientation error +51%, joint-position error +38%, joint-velocity error +81% — a profile that tracks key positions precisely while degrading orientation and joint-space fidelity.
Zero-Shot Onto Real Hardware
The nominal policy deploys zero-shot on a real Unitree G1 — no real-robot fine-tuning, no dedicated sim-to-real pipeline, no filtering of policy commands. Inference runs on an Intel Core Ultra 7 165H laptop over a direct cable link, reading state via ROS 2 and sending joint targets in a 50 Hz control loop identical to simulation. The robot tracks references spanning locomotion, crouching, dancing and loco-manipulation-style motions, and motions can be repeated or concatenated without resetting the controller.
Robustness held up under abuse: the G1 maintained or recovered balance while being manually pushed and dragged mid-motion, and kept tracking on a soft mattress — ground never seen in training. Hand-force randomization during training proved its value for forceful interaction tasks rather than for tracking accuracy per se, one of the study's cleaner dissociations.
Why the Field Needs Papers Like This
Motion-tracking papers typically showcase a single configuration's highlight reel. YAHMP's contribution is the opposite: a modular, open-source testbed where each design choice can be ablated in isolation, with honest accounting of what each knob costs and buys. For labs standing up humanoid tracking stacks in 2026 — and for the sim-to-real gap debate more broadly — it converts folklore into evidence.
