RT-1

Google, US · December 2022

Robotics at Google, Everyday Robots and Google Research

Action policyOpen weights, Apache-2.0Also written Robotics Transformer 1, RT-1 Robotics Transformer
Parameters
35 M
The paper states 35M total parameters.
GPU memory
0.1 GB
weights at bf16, computed. Fits an 8 GB card
Inference latency
not published
per action step
Weights
Apache-2.0
downloadable checkpoint

What it is

RT-1 is the original Robotics Transformer, a 35M parameter language-conditioned imitation policy that maps camera images and a task string to discretized action tokens. It has no vision-language backbone, which is exactly why it fits on the robot and runs closed loop at 3 Hz. Its evidence base is real hardware only, 130k episodes across more than 700 tasks gathered with 13 robots over 17 months. It is still the cheapest historically important checkpoint to run, but the 3 Hz rate and the absence of web-scale semantics put it well behind current VLAs on anything requiring open-vocabulary understanding.

Architecture

Backbone
No VLM backbone. FiLM conditioned EfficientNet image tokenizer, TokenLearner compression to 8 tokens per image, and a decoder-only Transformer with 8 self-attention layers.
Action head
Discretized action tokens over 11 dimensions (7 for the arm: x, y, z, roll, pitch, yaw, gripper opening; 3 for the mobile base: x, y, yaw; plus a mode switch).
Control rate
3 Hz
Parameters
The paper states 35M total parameters. The decoder-only Transformer accounts for 19M of these, the rest sits in the FiLM EfficientNet tokenizer. Only one model size was released.
Pretraining data
Over 130k real robot demonstration episodes covering over 700 tasks, collected with a fleet of 13 robots over 17 months. A bin-picking dataset from a Kuka arm was added for the multi-robot transfer study.
Embodiments
Everyday Robots mobile manipulator (7 DoF arm plus mobile base), Kuka arm (bin-picking data used only in the data-mixing transfer experiment)

What hardware it needs

This is the section most people came for, so it is worth being precise about which numbers are measured and which are arithmetic.

Computed from 35 M parameters. Weights only, so treat it as a floor: activations, image encoder intermediates and the CUDA context come on top, in practice 30 to 50 percent more for inference.
PrecisionWeightsNote
fp320.1 GBTraining master weights, rarely used for inference
bf16 or fp160.1 GBThe usual way these checkpoints are served
int80.0 GBQuantised, expect some loss of precision on fine motions
int40.0 GBAggressive quantisation, verify success rate before trusting it

Smallest real card that fits the bf16 weights plus a 40 percent runtime allowance: 8 GB. That is an estimate for inference, not for fine tuning. Full fine tuning also holds optimiser state and gradients, which typically costs several times the weights unless you use LoRA or a comparable adapter.

Published figures. These come from the model authors, not from this site.
Control rate3 Hz

Reported results

Grouped by what the evaluation actually asked. Values from different suites are not comparable with each other, so each one keeps its suite and split.

Simulation

No results in this category are published for this model.

Real world

Rollouts on physical hardware. The setups differ, so read these as evidence, not as a ranking.

  • RT-1 real-robot evaluation seen tasks, over 700 instructions
    97%
    success rateAuthors' own evaluation. Reported as 25 points above BC-Z and 32 points above Gato.source
  • RT-1 real-robot evaluation distractor robustness
    83%
    success ratesource
  • RT-1 real-robot evaluation unseen instructions
    76%
    success rateAuthors' own evaluation, 24 points above the next best baseline.source
  • SayCan long-horizon kitchen Kitchen1, up to 50 steps
    67%
    execution success ratesource
  • RT-1 real-robot evaluation background robustness
    59%
    success rateThe weakest of the four robustness categories, which is where RT-1 degrades first.source

Fine tuned tasks

No results in this category are published for this model.

On real hardware

Trained and evaluated entirely on real hardware. 13 robots, 17 months of data collection, over 700 instructions at 97 percent success on seen tasks and 76 percent on unseen instructions. Long-horizon kitchen sequences of up to 50 steps reached 67 percent execution success in Kitchen1.

Fine tuning it yourself

The released repository contains three SavedModel checkpoints (RT-1 trained on 700 tasks, RT-1 multi-robot on EDR plus Kuka, and RT-1 sim-real). The README covers loading and inference, not a supported fine-tuning pipeline, so adapting it to a new robot means rewriting the data pipeline to the RT-1 action discretization.

Where it helps, where it does not

Strengths

  • Genuinely small at 35M parameters, so it runs closed loop at 3 Hz without cloud inference
  • Weights and code are Apache-2.0 and the checkpoints ship inside the repository
  • The largest real-robot evidence base of its generation (130k episodes, 700 tasks, 17 months)
  • Documented positive transfer when mixing data from a second robot type (22 percent to 39 percent)

Limits

  • 3 Hz control is too slow for contact-rich or reactive manipulation
  • No vision-language backbone, so semantic generalization is far below later VLAs
  • Background robustness drops to 59 percent, the clearest failure mode
  • Single-arm plus mobile base only, and the hardware (Everyday Robots) no longer exists
  • The GitHub repository was archived on 7 April 2025 and is read-only, and the code is TensorFlow-era

Sources

Everything on this page was taken from these documents. Where they disagree with what you read here, they win.

Entry last checked 2026-08-11.

Try a policy on a real arm

A physical SO-100 is online and free to drive in the browser, and five of the models in this arena can be fine tuned on a dataset you record with your own.