VLA Gallery / All models / Humanoid VLA

LeVERB

LeVERB (Latent Vision-Language-Encoded Robot Behavior)

UC Berkeley·NTNU·Simon Fraser University·Carnegie Mellon University · 2025-06

VLAHumanoid VLASystem 2 trainable parameters 102.56Mlatent actions

Hierarchical VLA that links a vision-language layer and a whole-body control layer through a latent action vocabulary (latent verbs); trained only in sim and transferred zero-shot to G1.

Paper

LeVERB: Humanoid Whole-Body Control with Latent Vision-Language InstructionarXiv 2506.13751

Overview

Category

Humanoid VLA

Type

VLA

Released

2025-06

Organization

UC Berkeley·NTNU·Simon Fraser University·Carnegie Mellon University

Open weights

Unverified

Parameters

System 2 trainable parameters 102.56M; System 1 not disclosed (2 layers·128-dim)

Architecture

Backbone LLM

None (ViT-B vision·text encoder)

Vision encoder

ViT-B vision·text encoders (frozen)

Action head

Latent verbs → transformer student policy (latent actions)

Chunk

Not disclosed

Control rate

50 Hz

Architecture (original summary)

System 2 (LeVERB-VL): frozen ViT-Base vision·text encoders (768-dim each) → 256-dim latent 'verb', 10 Hz / System 1 (LeVERB-A): 2-layer·4-head·hidden 128 Transformer student, joint position actions 50 Hz

Training

Pretraining accelerators

Ada 6000 × 2 (System 2)

Pretraining duration

Not disclosed

Pretraining steps & batch

Steps not disclosed · global batch 512 (System 2)

GPU/TPU-hours

Not disclosed

Fine-tuning

N/A

Data type

Simulation·synthetic / motion capture

Data scale

Photoreal rollouts of 154 trajectories each randomized 100 times, 17.1 hours; language-only 2.7 hours (500 trajectories)

Main datasets

LeVERB-Bench (IsaacSim ray-traced rendering of retargeted motion capture)

Tasks & evaluation

LeVERB-Bench (10 categories, 150+ tasks); simple visual navigation 80%, overall 58.5%; 7.8× over a naive hierarchical whole-body VLA

Robot

Unitree G1 whole body (visual navigation·loco-manipulation)

Inference

Inference device & speed

System 2 10 Hz, System 1 50 Hz; hardware not disclosed

Inference memory

Not disclosed

Weights bf16 (calc.)

Not disclosed

Where it runs

Not disclosed

Reinforcement learning & base

Reinforcement learning

Partial (RL low-level controller) · System 2 trained by supervised learning; System 1 distills an RL teacher policy via DAgger

Base control

Humanoid legged (Commands/latents → RL locomotion controller) · LeVERB-VL latent verbs (10 Hz) → LeVERB-A directly outputs whole-body joint positions for walking·turning·sitting (visual navigation)

Notes

Key features

Hierarchical VLA that links a vision-language layer and a whole-body control layer through a latent action vocabulary (latent verbs); trained only in sim and transferred zero-shot to G1. System 1 distills a vision-language-agnostic RL teacher via DAgger.

Notes

Data hours cover only 17.1 h of photorealistic rollouts (language-only 2.7 h separate). Parameters (B) is null because total parameters are Not disclosed. Simulator·env·GPU for System 1 teacher RL Not disclosed (verification NOT FOUND); see separate LeVERB-A row. arXiv v3 2025-09-25.

Verification

Confirmed