Humanoid Native VLA

Abstract

Current Vision-Language-Action (VLA) systems generally infer robot actions using a single action-generation pipeline operating at a substantially uniform temporal resolution.

While suitable for many robotic manipulators, this architecture does not align with the heterogeneous dynamics of humanoid robots.

A humanoid simultaneously performs locomotion, manipulation, gaze control, balance, posture regulation, communication, and interaction.

These activities exhibit substantially different latency requirements.

This paper introduces Humanoid Native VLA, a reference architecture that distributes VLA inference across regional action generators while maintaining a shared semantic world model and synchronized whole-body state.

Regional inference frequencies may differ and may change dynamically during operation.

The resulting architecture enables efficient compute utilization, reduced latency, regional specialization, and closer integration with humanoid operating systems and TrustBoundary execution governance.

1. Introduction

Humanoid robots differ fundamentally from industrial manipulators.

Industrial robots typically optimize

1. one arm

2. fixed workspace

3. deterministic trajectories

Humanoids continuously coordinate

• arms

• legs

• torso

• neck

• eyes

• hands

• locomotion

• social interaction

Therefore,

a single global action decoder is unlikely to represent the optimal long-term architecture.

2. Humanoid Native VLA

The proposed architecture separates

semantic reasoning

from

regional action generation.

A shared semantic layer produces

1. task

2. world understanding

3. object relationships

4. intent

Regional inference systems specialize this information into body-specific actions.

3. Shared Runtime State

Every regional inference system accesses

• world model

• robot state

• task

• affordances

• capabilities

• resource reservations

• execution history

The runtime therefore behaves as one coherent humanoid.

4. Regional Action Heads

Rather than one action output,

multiple output heads generate

• hand actions

• locomotion

• balance

• torso

• gaze

• manipulation

• posture

Each may employ different neural architectures.

5. Multi-rate Inference

This paper proposes that inference frequency becomes a schedulable runtime resource.

Rather than fixing inference at one frequency,

each regional action generator may execute independently.

Example only

Region Relative Rate

Hand Highest

Foot High

Arm Medium

Torso Medium

Head Low

Face Lower

Language Lowest

The frequencies themselves are implementation-dependent.

The architectural contribution is regional independence.

6. Whole Body Synchronization

Regional independence does not imply independent robots.

Synchronization maintains

• stability

• collision avoidance

• momentum

• support

• temporal consistency

• shared execution state

7. TrustBoundary Integration

Unlike existing VLAs,

Humanoid Native VLA is designed around governed execution.

Regional outputs become

Proposed Actions

which are validated before execution.

TrustBoundary therefore becomes part of inference rather than an external safety filter.

8. Future Work

Future repository releases will include

• regional schedulers

• Isaac demonstrations

• G1 demonstrations

• TrustBoundary shell integration

• Humanoid OS API

• regional validator SDK

• body-region-local inference

• hardware abstraction

• distributed compute