HOVER

A versatile neural whole-body controller from NVIDIA and others where one policy works across many different command modes.

HOVER was released in October 2024 by NVIDIA (Linxi Fan, Yuke Zhu, and others) together with CMU, UC Berkeley, UT Austin, and UCSD, published at ICRA 2025. Navigation, loco-manipulation, and tabletop manipulation each need a different control interface for a humanoid: navigation cares about torso (root) velocity, while tabletop manipulation cares about upper-body joint positions. Previously, each interface was usually trained as its own separate, non-interchangeable policy. HOVER first trains a teacher whole-body motion-tracking policy that imitates AMASS human motion data in simulation, then distills it into a student policy; during training, mode masks and sparse masks are applied separately to upper- and lower-body commands, letting the same policy switch among the command modes used by H2O, OmniH2O, ExBody, and HumanPlus with no retraining needed. The code is open-sourced on Isaac Lab, and it was deployed on a real Unitree H1.

ExampleThe same HOVER policy can either take just torso-velocity commands to make the H1 walk, or switch to a VR teleoperation mode that tracks only the head and hand positions.

Also called

Versatile Neural Whole-Body Controller for Humanoid Robots

Related

Whole-Body Control · Learning-Based Whole-Body Control · Policy Distillation · Motion Tracking · OmniH2O · NVIDIA Generalist Embodied Agent Research Lab

Sources

HOVER: Versatile Neural Whole-Body Controller for Humanoid Robots (arXiv 2410.21229)
HOVER project page
NVlabs/HOVER (GitHub)

As of

2025-03