VLA Gallery / All models / Industry models

SARA-RT

Google DeepMind · 2023-12

기타Industry modelsNot discloseddiscrete tokens

An efficiency technique that replaces quadratic-complexity attention in pretrained·fine-tuned transformer robot policies with linear attention.

Paper

SARA-RT: Scaling up Robotics Transformers with Self-Adaptive Robust AttentionarXiv 2312.01990

Overview

Category

Industry models

Type

기타

Released

2023-12

Organization

Google DeepMind

Open weights

No

Parameters

Not disclosed

Architecture

Backbone LLM

RT-2 backbone (UL2, converted to linear attention)

Vision encoder

—

Action head

RT-2 action tokens (discrete tokens)

Chunk

Not disclosed

Control rate

Not disclosed

Architecture (original summary)

Converts attention of existing transformer policies (RT-2, Point Cloud Transformer) to linear attention via up-training

Training

Pretraining accelerators

Not disclosed

Pretraining duration

Not disclosed

Pretraining steps & batch

Not disclosed

GPU/TPU-hours

Not disclosed

Fine-tuning

Not disclosed

Data type

N/A

Data scale

N/A

Main datasets

N/A

Tasks & evaluation

Accuracy +10.6%, speed +14% vs RT-2 (blog)

Robot

Not disclosed (applied to RT-2·Point Cloud Transformer policies)

Inference

Inference device & speed

14% faster than RT-2 (with short image history input); hardware not disclosed

Inference memory

Not disclosed

Weights bf16 (calc.)

Not disclosed

Where it runs

Not disclosed

Reinforcement learning & base

Reinforcement learning

Not used · linear attention up-training (supervised learning)

Base control

Fixed · No base control (RT-2 action space: EE displacement·gripper·terminate; PCT is Kuka arm grasping)

Notes

Key features

An efficiency technique that replaces quadratic-complexity attention in pretrained·fine-tuned transformer robot policies with linear attention. Reported 10.6% higher accuracy and 14% faster than RT-2 with short image-history input.

Notes

An efficiency technique (not a standalone model), so type is Other. arXiv 2312.01990 (2023-12), blog 2024-01, so by the rules it is dated 2023-12 and falls outside the 2024~2025 range. RT-Trajectory from the same blog: 63% on 41 new tasks vs RT-2 29%.

Verification

Not verified