OpenHLM An Empirical Recipe for Whole-Body
Humanoid Loco-Manipulation

Yingdong Hu 1,2,3 * Haodong Zhu 1 * Boyuan Zheng 1,2 * Yihang Hu 1,2 * Tong Zhang 1,2,3 *

Zunhao Chen 1 Junming Zhao 1 Ruiqian Nai 1 Yang Gao 1,2,3 †

1 Tsinghua University 2 Shanghai Qi Zhi Institute 3 Spirit AI

* Core contributors Corresponding author

A humanoid in the wild

OpenHLM is a recipe for whole-body humanoid loco-manipulation: it maps a language instruction and the robot's observation images into coordinated motion across every one of the robot's degrees of freedom.

So what can it do? Mix in just 20 outdoor demos with our lab data, and the humanoid moonlights as the campus janitor—fully autonomous, outdoors, no mocap. As far as we know, no humanoid has done this in the wild before.

autonomous 2x

Comparison with SOTA

How does OpenHLM stack up against other humanoid VLAs? We compare at the system level with two representative baselines: GR00T N1.6 and Ψ₀.

The evaluation is a challenging long-horizon task: the robot walks between a low table, a medium table, and a tall shelf, picking two language-specified fruits—one per hand—and placing them into separate containers. This spans the humanoid's full vertical workspace and requires language grounding across 20 ordered fruit pairs. OpenHLM reaches 87.5% task progress using less than half the demonstration time of either baseline.

autonomous 2x
1.14 h demo time
Task Progress 87.5%
autonomous 2x

GR00T N1.6

2.70 h demo time
Task Progress 57.5%
autonomous 2x

Ψ₀

2.70 h demo time
Task Progress 48.8%

HLM Tasks

Want more tasks? We introduce 12 language-conditioned tasks that target different aspects of whole-body loco-manipulation behavior, organized into four categories: (1) pick-and-place with locomotion, (2) whole-body workspace extension, (3) using body parts as manipulators, and (4) loco-manipulation under environmental constraints. OpenHLM achieves strong performance across all 12 tasks, with average task progress exceeding 90%.

Pick & Place autonomous 2x

Cola Placement

01 / 12

Workspace Extension autonomous 2x

Shelf Cup Transfer

02 / 12

Body-as-Tool autonomous 2x

Bottle Disposal

03 / 12

Constraints autonomous 2x

Jar Opening

04 / 12

Body-as-Tool autonomous 2x

Toy Stowing

05 / 12

Constraints autonomous 2x

Sword Extraction

06 / 12

Constraints autonomous 2x

Cart Pushing

07 / 12

Constraints autonomous 2x

Shuttlecock Setup

08 / 12

Pick & Place autonomous 2x

Pig Placement

09 / 12

Pick & Place autonomous 2x

Gum Can Placement

10 / 12

Workspace Extension autonomous 2x

Shelf Cube Transfer

11 / 12

Constraints autonomous 2x

Pouring

12 / 12

How did we get here?

None of this comes from a single trick. We construct OpenHLM through controlled experiments, one design decision at a time, in three phases that build up to a concrete recipe for whole-body humanoid loco-manipulation.

Low-level controller and teleoperation: how to design the controller and its teleop interface for high-quality whole-body demonstrations. Whole-body VLA policy design: which adaptations turn a VLA built for static and wheeled robots into a whole-body humanoid policy. Heterogeneous co-training: whether cheaper data sources can extend the policy beyond what whole-body teleop alone covers.

OpenHLM teaser figure

Low-level controller and teleoperation

We follow a two-level hierarchical control framework: a high-level policy (the operator during data collection, the learned VLA at deployment) emits reference whole-body commands, and a lightweight low-level controller tracks them. With this framework fixed, the interface between the two determines both what the operator can express and what action space the VLA learns.

We compare three teleoperation methods representative of recent humanoid systems: decoupled control (dual-arm IK plus an RL lower-body controller tracking a base-velocity and root-height command, 21-D), VR 3-point (head and wrist poses plus a navigation command, 24-D), and joint-based whole-body teleoperation (a portable motion-capture rig retargeted in real time to every humanoid joint, 32-D). We collect matched demonstrations under each interface and train one VLA per method.

Teleop Method Comparison

Select a task to compare task progress, rollout time, and footsteps. N/A marks tasks the interface cannot express by construction.

Method
Prog. Task progress ↑
Time Rollout duration ↓
Steps Footsteps per rollout ↓
Decoupled Control
VR 3-Point Teleoperation
Joint-Based Whole-Body Teleop. (Ours)
21-D action Cola Placement
autonomous 2x

Decoupled Control

24-D action Cola Placement
autonomous 2x

VR 3-Point

32-D action Cola Placement
autonomous 2x

Joint-Based Whole-Body

The numbers tell a clear story. Joint-based whole-body teleoperation is the only interface that completes all three tasks. The two alternatives expose only a subset of the humanoid's degrees of freedom, so tasks like a foot depressing a pedal or the whole body squatting under a shelf are unreachable by construction. Based on these results, we adopt joint-based whole-body teleoperation as the data-collection interface that all later phases build on.

Whole-body VLA policy design

Existing VLAs bring vision-language reasoning and manipulation priors, but nearly all target static or wheeled dual-arm platforms—none were designed for humanoid loco-manipulation. So how do we adapt one into a whole-body humanoid policy, and which design choices actually matter? We organize the exploration into three families: the action and proprioception interface, the role of pretraining, and faster action generation—ablating one component at a time on a 4-task subset.

91.3%
Ours (Default)
82.1%
Random Action Projection
87.1%
Humanoid-Native Ordering
86.2%
Relative Actions
81.2%
No Proprioception
59.6%
PaliGemma Init
41.7%
Random Init
72.9%
One-Step Flow
70.8%
Drifting Model
Interface Adaptations Barely Matter

Adapting π₀.₅'s action space to the humanoid involves four choices: projection initialization, action ordering, absolute vs. relative targets, and proprioceptive input. Flipping any one produces only a slight drop—no single choice is the bottleneck.

Non-Humanoid Pretraining Transfers

Initialized from π₀.₅ (pretrained on dual-arm robots), the policy reaches 91% task progress; PaliGemma (same architecture, vision-language only) drops to 60%; random init collapses to 42%. The cross-embodiment gap is real, but dwarfed by the gap between any robot pretraining and none.

Multi-Step Inference Wins

Counterintuitively, both one-step alternatives reach lower validation action MSE than π₀.₅'s 10-step flow matching—yet on the robot they underperform it by ~20 task-progress points. So we keep multi-step inference.

Heterogeneous co-training

With the whole-body VLA established, scaling to every task through loco-manipulation teleop is expensive. We turn to two cheaper data sources and ask whether co-training can incorporate them effectively: stationary teleoperation (the same humanoid with feet planted, manipulation only) and HuMI (robot-free demonstrations captured with low-cost wearable devices). The question is whether either can extend the policy to held-out tasks that whole-body teleop never covers.

Whole-Body Teleop illustration

Whole-Body Teleop

21 min / held-out task

Full loco-manipulation teleop on the humanoid; the expensive oracle data source.

Stationary Teleop illustration

Stationary Teleop

13 min / held-out task

The same humanoid with feet planted: manipulation only, no locomotion—cheaper to collect.

HuMI illustration

HuMI

7 min / held-out task

A robot-free rig of hand-held UMI grippers and body-pose trackers; the humanoid stays out of the loop.

Task Progress
on Held-Out Tasks 9–11

With the policy trained on whole-body teleop for Tasks 1–8, we co-train it with cheaper data on three held-out tasks (9–11) that whole-body teleop never covers. Both streams lift average progress on these tasks from 36% to near the oracle—with no extra whole-body teleop.

8-Task Baseline 36%

Trained only on Tasks 1–8, never on 9–11.

HuMI Co-Training 84%

Adds robot-free HuMI demos for Tasks 9–11.

Stationary Co-Training 89%

Adds feet-planted teleop demos for Tasks 9–11.

11-Task Teleop Oracle 96%

Full whole-body teleop on all tasks—the expensive upper bound.