AY-ROBOTSAI Teleoperation
Sign Up
OXE

RT-1-X

Open X-Embodiment, US · October 2023

Open X-Embodiment Collaboration

Action policyOpen weights, Apache-2.0Also written RT-X, Open X-Embodiment RT-1-X
Parameters
not published
RT-1-X reuses the RT-1 architecture, for which the RT-1 paper reports 35M parameters.
GPU memory
not published
weights at bf16, computed
Inference latency
not published
per action step
Weights
Apache-2.0
Apache-2.0 for the software, CC-BY 4.0 for the other materials

What it is

RT-1-X is the RT-1 architecture retrained on the pooled Open X-Embodiment mixture of 60 datasets from 34 labs, 22 embodiments and over 1M trajectories. Its contribution is empirical evidence for positive cross-embodiment transfer, worth 50 percent over training on a single dataset in the small-data regime. It is the only RT-X model whose weights were actually released, under Apache-2.0 code and CC-BY materials, with TensorFlow and JAX checkpoints. Because it inherits RT-1 and not a VLM, it does not give you open-vocabulary semantics, and the published documentation omits parameter count, control rate and latency.

Architecture

Backbone
The RT-1 architecture (FiLM EfficientNet tokenizer, TokenLearner, decoder-only Transformer) retrained on the pooled Open X-Embodiment mixture. No VLM backbone.
Action head
Discretized action tokens over a normalized 7-dimensional end-effector action space (x, y, z, roll, pitch, yaw, gripper opening) shared across all embodiments.
Parameters
RT-1-X reuses the RT-1 architecture, for which the RT-1 paper reports 35M parameters. The Open X-Embodiment paper does not restate a parameter count for RT-1-X, so no figure is asserted here.
Pretraining data
Open X-Embodiment: 60 existing robot datasets from 34 robotics research labs, covering 22 robot embodiments, more than 1M real robot trajectories, 527 skills and 160,266 tasks.
Embodiments
22 robot embodiments pooled across 34 labs, including single-arm manipulators, bi-manual setups and quadrupeds

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.

The authors publish no memory or latency figure for this model. Everything above is computed from the parameter count.

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.

  • Open X-Embodiment cross-lab evaluation small-data domains, in-distribution skills across 6 academic labs
    50%
    improvement over RT-1 or the Original Method trained on the single datasetAuthors' own evaluation. The gain is specific to the small-data regime and is not claimed for data-rich domains.source

Fine tuned tasks

No results in this category are published for this model.

On real hardware

Evaluated on real hardware at six academic labs on in-distribution skills. The core real-world result is positive transfer: pooling 60 datasets lifts performance on small datasets by 50 percent over a policy trained on that dataset alone.

Fine tuning it yourself

The repository provides an inference Colab and rt1_inference_example.py showing how to load the checkpoint and compare predicted against ground-truth actions. There is no first-party fine-tuning recipe, so adapting to a new robot means mapping your data into the RT-1-X normalized 7-DoF action space and the RLDS dataset format used by Open X-Embodiment.

Where it helps, where it does not

Strengths

  • One of very few genuinely open cross-embodiment checkpoints, released in both TensorFlow and JAX form
  • Documented positive transfer of 50 percent in the small-data regime, which is the regime most labs are in
  • The underlying Open X-Embodiment dataset is public and has become the standard pretraining mixture
  • Permissive licensing (Apache-2.0 code, CC-BY materials)

Limits

  • No VLM backbone, so open-vocabulary and semantic generalization stay at RT-1 level
  • The 50 percent gain is reported only for small-data domains, not as a universal improvement
  • The shared 7-DoF end-effector action space forces every embodiment into one representation and excludes richer morphologies
  • No parameter count, control rate or latency is published for RT-1-X itself
  • Checkpoints are distributed via a Google Cloud Storage bucket rather than Hugging Face, and the inference path is a Colab plus a minimal example script

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.

Download the clientDrive a real arm now