VLA Gallery / All models / Humanoid whole-body controllers

LeVERB-A

LeVERB-A (LeVERB System 1)

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

VAHumanoid whole-body controllersNot disclosedRL policy

Distills vision-language-agnostic RL teachers into a transformer student via DAgger, converting latent commands into whole-body actions.

Paper

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

Overview

Category

Humanoid whole-body controllers

Type

VA

Released

2025-06

Organization

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

Open weights

Unverified

Parameters

Not disclosed (2 layers·128-dim)

Architecture

Backbone LLM

None

Vision encoder

—

Action head

Latent verbs → joint positions (RL policy)

Chunk

Not disclosed

Control rate

50 Hz

Architecture (original summary)

2-layer · 4-head · hidden 128 Transformer student; LeVERB-VL's 256-dim latent verbs → joint position actions 50 Hz

Training

Pretraining accelerators

Not disclosed

Pretraining duration

Not disclosed

Pretraining steps & batch

Not disclosed

GPU/TPU-hours

Not disclosed

Fine-tuning

N/A

Data type

Simulation·synthetic / motion capture

Data scale

Not disclosed (uses LeVERB-Bench based on retargeted motion capture)

Main datasets

LeVERB-Bench

Tasks & evaluation

Evaluated as the full LeVERB system (simple visual navigation 80%, overall 58.5%)

Robot

Unitree G1

Inference

Inference device & speed

50 Hz; hardware not disclosed

Inference memory

Not disclosed

Weights bf16 (calc.)

Not disclosed

Where it runs

Not disclosed

Reinforcement learning & base

Reinforcement learning

Used (RL controller training) · RL teacher policy distilled into a Transformer student via DAgger

Base control

Humanoid legged (RL policy drives joints directly) · Takes latent verbs as input → directly outputs whole-body joint position actions (50 Hz) for walking and sitting (distilled from RL teacher policy)

Notes

Key features

Distills vision-language-agnostic RL teachers into a transformer student via DAgger, converting latent commands into whole-body actions. A transformer whole-body controller placed below a VLA.

Notes

Simulator·env count·GPU for teacher RL Not disclosed (verification NOT FOUND). Same paper as the LeVERB row.

Verification

Confirmed