VLA Gallery / All models / Reference baselines (pre-2024)

RT-1

Robotics at Google; Everyday Robots; Google Research(Brain Team) · 2022-12

VAReference baselines (pre-2024)35Mdiscrete tokens

A 35M real-robot policy conditioned on language instructions via FiLM, without a pretrained VLM.

Paper

RT-1: Robotics Transformer for Real-World Control at ScalearXiv 2212.06817

Overview

Category

Reference baselines (pre-2024)

Type

VA

Released

2022-12

Organization

Robotics at Google; Everyday Robots; Google Research(Brain Team)

Open weights

Yes

Parameters

35M (tokenizer 16M + Transformer 19M; TokenLearner not stated separately)

Architecture

Backbone LLM

None

Vision encoder

EfficientNet-B3 + TokenLearner (16M)

Action head

256-bin discrete tokens (discrete tokens)

Chunk

None (1 step)

Control rate

3 Hz

Architecture (original summary)

FiLM EfficientNet-B3 tokenizer (16M) + TokenLearner + Transformer (8 self-attention layers, 19M); 256-bin discrete token per action dimension (arm 7 · base 3 · mode), no chunking, 3 Hz

Training

Pretraining accelerators

Not disclosed

Pretraining duration

Not disclosed

Pretraining steps & batch

Not disclosed

GPU/TPU-hours

Not disclosed

Fine-tuning

N/A (single-stage training)

Data type

Real-robot teleoperation

Data scale

About 130k demos (13 robots, 17 months); data-mixture experiments add 518k successful simulation trajectories and 209k Kuka QT-Opt episodes

Main datasets

Self-collected RT-1 data; mixing experiments: simulation, Kuka QT-Opt

Tasks & evaluation

3,000+ real-robot trials; 200+ training tasks, 744 instructions; 21 unseen instructions, 53 held-out tasks, 15 long-horizon instructions

Robot

13 mobile manipulators (office kitchen; arm 7 variables·base 3 variables)

Inference

Inference device & speed

3 Hz control, inference budget under 100 ms; inference hardware not disclosed

Inference memory

Not disclosed

Weights bf16 (calc.)

0.1 GB

Where it runs

Not disclosed

Reinforcement learning & base

Reinforcement learning

Not used · Imitation learning (behavior cloning). Data-mixing experiments added 209k episodes collected on Kuka robots by the RL policy QT-Opt

Base control

Wheeled (Direct base velocity/pose output) · Policy directly outputs base x, y, yaw motion as discrete tokens with 256 bins (mode tokens switch between arm, base, and termination)

Notes

Key features

A 35M real-robot policy conditioned on language instructions via FiLM, without a pretrained VLM. With discrete action tokens and 17 months of teleoperation data from 13 robots, it became the baseline for the later RT family.

Notes

Notes classify it as VA (language-conditioned but no pretrained VLM). Training compute is absent from all primary sources, so secondary-source compute figures are treated as unsupported. Checkpoints rt1main·rt1multirobot·rt1simreal released (Apache).

Verification

Confirmed