OXE

RT-2-X

Open X-Embodiment, US · October 2023

Open X-Embodiment Collaboration

Foundation VLAClosed weightsAlso written RT-X, Open X-Embodiment RT-2-X
Parameters
55 B
The project page describes RT-2-X (55B) as one of the biggest models to date.
GPU memory
102 GB
weights at bf16, computed
Inference latency
not published
per action step
Weights
closed
no public checkpoint

What it is

RT-2-X is RT-2 co-fine-tuned on the Open X-Embodiment mixture instead of a single robot dataset. Its result matters conceptually: pooling data from 22 embodiments made the 55B model 3x better on emergent skills and produced behaviors RT-2 could not perform at all. Practically it is unavailable, since the collaboration released RT-1-X checkpoints but not RT-2-X. No control frequency, latency or memory figures were published for it. Use it as evidence for cross-embodiment pretraining, not as a candidate model.

Architecture

Backbone
RT-2 (PaLI-X based vision-language model) co-fine-tuned on the Open X-Embodiment mixture in addition to web vision-language data.
Action head
Actions as text tokens, inherited from RT-2, over the shared 7-dimensional end-effector action space (x, y, z, roll, pitch, yaw, gripper opening).
Parameters
The project page describes RT-2-X (55B) as one of the biggest models to date. Unlike RT-1-X, no RT-2-X checkpoint was released.
Pretraining data
Web-scale vision-language data from the PaLI-X pretraining plus Open X-Embodiment (60 datasets, 34 labs, 22 embodiments, over 1M trajectories, 527 skills, 160,266 tasks).
Embodiments
Multiple embodiments from the Open X-Embodiment collection, including the Everyday Robots mobile manipulator and academic lab arms

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 55 B 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
fp32205 GBTraining master weights, rarely used for inference
bf16 or fp16102 GBThe usual way these checkpoints are served
int851 GBQuantised, expect some loss of precision on fine motions
int426 GBAggressive quantisation, verify success rate before trusting it

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.

  • RT-2 emergent skills evaluation emergent skills not present in the robot training data
    3 x
    improvement over RT-2Authors' own evaluation. The project page states RT-2-X demonstrates skills the RT-2 model was not previously capable of.source

Fine tuned tasks

No results in this category are published for this model.

On real hardware

Evaluated on real hardware within the Open X-Embodiment collaboration. The headline real-world claim is that cross-embodiment co-training unlocks skills RT-2 could not do at all, quantified as 3x on emergent skill evaluations.

Fine tuning it yourself

Not possible. The Open X-Embodiment repository releases RT-1-X checkpoints only; RT-2-X weights are not part of the release.

Where it helps, where it does not

Strengths

  • Shows that cross-embodiment data adds capability to a large VLA, not just robustness
  • 3x over RT-2 on emergent skills is the strongest transfer result in the RT-X paper
  • Benefits from both web-scale semantics and 22 embodiments of robot data

Limits

  • Weights were never released, only RT-1-X was published, so RT-2-X is not deployable
  • At 55B it inherits RT-2's cloud-only serving profile
  • No published control rate, latency, or VRAM figures
  • No public fine-tuning path of any kind

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.