VLA Gallery / All models / Mobile manipulation

MoRE

Zhejiang University·Westlake University(MiLAB)·Monash University·HKUST(GZ) · 2025-03

VLAMobile manipulation9.82Bdiscrete tokens

A quadruped VLA that inserts many LoRA experts into Fuyu-8B as a sparse MoE, treats the VLA as a Q-function, and trains it with offline RL on auto-collected data mixing successes and failures.

Paper

MoRE: Unlocking Scalability in Reinforcement Learning for Quadruped Vision-Language-Action ModelsarXiv 2503.08007

Overview

Category

Mobile manipulation

Type

VLA

Released

2025-03

Organization

Zhejiang University·Westlake University(MiLAB)·Monash University·HKUST(GZ)

Open weights

No

Parameters

9.82B (Fuyu-8B + MoE-LoRA)

Architecture

Backbone LLM

Fuyu-8B + MoE-LoRA experts

Vision encoder

Fuyu-8B patch input (no separate vision encoder)

Action head

Autoregressive discrete tokens (discrete Q-function) → 12-D commands (discrete tokens)

Chunk

Not disclosed

Control rate

Not disclosed

Architecture (original summary)

Fuyu-8B (32 layers) with each layer's FFN replaced by a shared FFN + LoRA expert MoE (Top-K router), LoRA on attention → autoregressive discrete action tokens (discrete Q-function) → 12-dim discrete command → Walk These Ways RL controller

Training

Pretraining accelerators

8 × A100

Pretraining duration

About 100 h (expert data) / about 125 h (mixed-quality data)

Pretraining steps & batch

3 epochs · batch 16 per GPU

GPU/TPU-hours

800 (Converted · 8× A100 × 100 hours (expert data training; the 125-hour mixed-quality run does not restate GPU count))

Fine-tuning

Additional fine-tuning on a small amount of real-robot data (data scale·resources·time not disclosed)

Data type

Simulation·synthetic

Data scale

QUARD expert 1,822,405 VLA sets (6 tasks) + QUARD-Auto 440,732 sets (failure data, 4 tasks) + small real-robot set (size not stated)

Main datasets

QUARD; QUARD-Auto

Tasks & evaluation

QUARD 6-skill sim average 0.60 (QUART 0.44): Distinguish 0.82, Go to 0.80, Go avoid 0.59, Go through 0.57, Crawl 0.49, Unload 0.33; real robot only qualitative results (go to the computer, etc.)

Robot

Unitree Go2 quadruped (12 joints) + front RealSense D435; sim IsaacGym

Inference

Inference device & speed

Not disclosed

Inference memory

Not disclosed

Weights bf16 (calc.)

19.6 GB

Where it runs

Not disclosed

Reinforcement learning & base

Reinforcement learning

Used (VLA RL post-training) · A Fuyu-8B-based VLA (LoRA experts) trained with offline RL as a discrete Q-function (Bellman updates · conservative term · sparse reward of 1 on success); low level is the Walk These Ways RL policy

RL evidence

https://arxiv.org/abs/2503.08007

Base control

Quadruped (Commands/latents → RL locomotion controller) · VLA outputs 12-dim discrete command tokens → Walk These Ways RL policy controls joints

Notes

Key features

A quadruped VLA that inserts many LoRA experts into Fuyu-8B as a sparse MoE, treats the VLA as a Q-function, and trains it with offline RL on auto-collected data mixing successes and failures. Its average sim success over 6 skills was 0.60, ahead of baselines such as QUART (0.44), and adding failure data improved performance through RL.

Notes

Number of experts, LoRA rank, and Top-K value not stated; descriptions of 'frozen backbone, LoRA-only training' and of converting FFN→MoE are mixed; sparse reward (1 on success) + conservative regularization (α=0.5); inference speed not stated; no public code or weights link. Based on: GeRM: A Generalist Robotic Model with Mixture-of-experts for Quadruped Robot (arXiv 2403.13358) — reuses GeRM's Q-function training loss and QUARD-Auto data as is (quadruped MoE model from the same group)

Verification

Confirmed