HOVER: Versatile Neural Whole-Body Controller for Humanoid Robots

Tairan He, Wenli Xiao, Toru Lin, Zhengyi Luo, Zhenjia Xu, Zhenyu Jiang, Jan Kautz, Changliu Liu, Guanya Shi, Xiaolong Wang, Linxi Fan, Yuke Zhu

Seminal October 2024

Real H1 teleoperated from a VR headset while hand targets are masked on and off.

TL;DR One humanoid policy that accepts any mix of keypoint, joint-angle and root commands, distilled from a full-body motion-imitation teacher under random command masks. On a Unitree H1 it beats the single-mode specialists it replaces, for example in 11 of 12 real-world tracking metrics.

What they do

Learned humanoid controllers each fix one command interface: root velocity for locomotion, joint angles for expressive motion, or a few body keypoints for teleoperation. HOVER trains one policy for a 19-DoF Unitree H1 that accepts any masked subset of a shared command space, which covers the command spaces of ExBody, H2O, OmniH2O and HumanPlus and over 15 modes in total. Modes can be switched while the robot is moving.

How it works

Command space. Upper and lower body each take kinematic keypoint positions, joint angles and root targets (velocity, height, roll/pitch/yaw). Binary masks choose which entries are active.
Oracle teacher. A PPO policy with privileged full rigid-body state imitates AMASS human motion retargeted to the H1, with domain randomization for sim-to-real.
Masked distillation. The student sees 25 steps of proprioception plus masked commands. Each mode and sparsity mask bit is drawn from Bernoulli(0.5) at the start of each episode, and the student is trained with DAgger against the teacher’s actions.

Results

In simulation, HOVER beats each prior-work specialist on at least 7 of 12 metrics, e.g. global body position error 185 vs 275 mm in ExBody mode and 128 vs 149 mm in OmniH2O mode.
It also beats RL specialists for left-hand, right-hand, two-hand and head modes (right hand: 128 vs 220 mm global error).
A multi-mode policy trained with RL from scratch, using the same masking, has higher error on all 32 metric/mode pairs.
On the real H1, tracking 20 standing motions, HOVER wins 11 of 12 metrics (HumanPlus mode: 47.4 vs 51.0 mm). Mode switches during walking and Vision Pro teleoperation with randomly masked head and hands are shown qualitatively.

Limitations

The operator picks the mode; automatic mode switching is left as future work. Real-world numbers cover only 20 standing motions, and locomotion is shown only qualitatively. Specialists still win some metrics, e.g. ExBody’s root velocity error (0.428 vs 0.452 m/s).

Videos

Real H1 following an operator's VR motion with different keypoints tracked.

Simulated humanoids following root velocity commands and keypoint targets.

Figures

From Figure 1: One command space covering kinematic position tracking (blue), local joint angle tracking (yellow) and... — From Figure 1: One command space covering kinematic position tracking (blue), local joint angle tracking (yellow) and root tracking (purple), driven from VR, RGB, exoskeleton, MoCap or joystick interfaces. Masks select which commands are active.

From Figure 2: The HOVER student is distilled from the oracle policy with DAgger. Mode and sparsity masks, applied se... — From Figure 2: The HOVER student is distilled from the oracle policy with DAgger. Mode and sparsity masks, applied separately to upper and lower body, decide which commands the student sees.

From Figure 3: Prior-work specialists (blue) versus the HOVER generalist (green) in ExBody, HumanPlus, H2O and OmniH2... — From Figure 3: Prior-work specialists (blue) versus the HOVER generalist (green) in ExBody, HumanPlus, H2O and OmniH2O modes, on joint, body position, root velocity and root rotation errors.

From Figure 6: On the real robot, command modes switch during forward walking (ExBody to H2O) and during turning and ... — From Figure 6: On the real robot, command modes switch during forward walking (ExBody to H2O) and during turning and backward walking (HumanPlus to OmniH2O), and VR teleoperation switches between two-hand, head and left-hand modes.

Abstract

Humanoid whole-body control requires adapting to diverse tasks such as navigation, loco-manipulation, and tabletop manipulation, each demanding a different mode of control. For example, navigation relies on root velocity tracking, while tabletop manipulation prioritizes upper-body joint angle tracking. Existing approaches typically train individual policies tailored to a specific command space, limiting their transferability across modes. We present the key insight that full-body kinematic motion imitation can serve as a common abstraction for all these tasks and provide general-purpose motor skills for learning multiple modes of whole-body control. Building on this, we propose HOVER (Humanoid Versatile Controller), a multi-mode policy distillation framework that consolidates diverse control modes into a unified policy. HOVER enables seamless transitions between control modes while preserving the distinct advantages of each, offering a robust and scalable solution for humanoid control across a wide range of modes. By eliminating the need for policy retraining for each control mode, our approach improves efficiency and flexibility for future humanoid applications.

Summaries are written automatically from each paper's text. Figures and videos come from the paper and its project page; all credit goes to the authors.

Matt Strong