Open Source RoboticsRobots, software & learning

RT-1 / RT-1-X (Robotics Transformer)

Robotics model

FiLM-EfficientNet + TokenLearner + 8-layer Transformer VLA policy family from Google Research and Google DeepMind, spanning the archived Apache-2.0 RT-1 TensorFlow repo (3 SavedModel checkpoints) and the Open X-Embodiment RT-1-X TensorFlow and JAX checkpoints. [1] [2] [3] [4]

Setup & usage →CheckpointsModel hub / homepageGetting started

For
Robotics and VLA researchers; Open X-Embodiment benchmark practitioners; TensorFlow / JAX robot learning engineers [1] [3]
Parameters
Compact 8-layer transformer (~0.78 GB RT-1-X TF zip [1] [4]
Embodiments
Everyday Robots (EDR) mobile manipulator (RT-1), Kuka arm (rt1multirobot), Cross-embodiment 7-DoF arms in Open… [1] [3] [6]
Source rights
Named licenses are documented by part. Review licenses

Source researched · Not tested by us
Sources checked · Research details

Licenses & related hardware

Requirements · Research details and dates

On this page

Overview

RT-1 (google-research/robotics_transformer, commit 4569641b8111, archived read-only) and its cross-embodiment Open X-Embodiment successor RT-1-X (google-deepmind/open_x_embodiment, commit 9eeb68b989ef) form a foundational Vision-Language-Action (VLA) policy family that maps monocular workspace RGB frames and natural language task strings directly to discretized 7-DoF gripper actions. Five ungated checkpoints are hosted across Google Cloud Storage (RT-1-X TensorFlow and JAX) and Git LFS (three original RT-1 TensorFlow SavedModels). [1] [3] [4] [5] [6]

Model card summary

Architecture

FiLM-conditioned EfficientNet image tokenizer + TokenLearner token compression + 8-layer decoder-only Transformer (configs/transformer_mixin.gin: num_layers 8, num_heads 8, layer_size 128, token_embedding_size 512, feed_forward_size 512, vocab_size 256, time_sequence_length 6).

Verified directly from transformer_network.py and configs/transformer_mixin.gin.

[1]

Base model

EfficientNet (FiLM-conditioned image tokenizer) + Universal Sentence Encoder / language embedding

[1] [5]

Parameter count

Compact 8-layer transformer (~0.78 GB RT-1-X TF zip; ~1.37 GB JAX checkpoint; ~0.14 GB RT-1 main variables; exact scalar count not printed in README)

Exact parameter count is not stated in robotics_transformer README.md or open_x_embodiment README.md; verified checkpoint byte sizes are reported instead.

[1] [4]

Input modalities
rgblanguage

Single workspace RGB camera image history (sequence length 6) plus natural language task string; no wrist camera, depth, or proprioception.

[1] [3]

Output modalities
ee-pose-deltagripper-command

7 discretized gripper movement variables (x, y, z, roll, pitch, yaw, gripper opening) quantized into 256 bins (vocab_size 256); per OXE README each variable can represent absolute, delta, or velocity depending on embodiment convention.

[3]

Action space

7-DoF gripper control (x, y, z, roll, pitch, yaw, gripper opening) discretized into 256 token bins per dimension (vocab_size: 256) plus termination/mode tokens.

[1] [3]

Control frequency

3 Hz (333 ms per step documented in Open X-Embodiment README)

[3]

Trained / supported embodiments
  • Everyday Robots (EDR) mobile manipulator (RT-1)
  • Kuka arm (rt1multirobot)
  • Cross-embodiment 7-DoF arms in Open X-Embodiment (RT-1-X)

[1] [3] [6]

Training datasets
  • Open X-Embodiment multi-robot dataset mixture (RT-1-X)
  • Google Everyday Robots 700-task dataset, Kuka dataset, and sim+real data (RT-1 checkpoints)

[1] [3] [6]

Interfaces
  • TensorFlow SavedModel
  • JAX / Flax checkpoint loader (rt1_inference_example.py)
  • TF-Agents TransformerNetwork Python API

[1] [3]

Checkpoints

CheckpointFormatSizeAccessPinned revision
RT-1-X TensorFlow SavedModel (2,272,480 steps)other0.78 GBopen2023-10-10
RT-1-X JAX / Flax Checkpoint (GCS bucket)other1.37 GBopenmain
RT-1 Main (700 tasks, Git LFS SavedModel)other0.14 GBopen4569641b8111
RT-1 Multi-Robot (EDR + Kuka, Git LFS SavedModel)other0.45 GBopen4569641b8111
RT-1 Sim + Real (Git LFS SavedModel)other0.14 GBopen4569641b8111

Verified GCS HTTP 200 (0.775 GB zip, 1.368 GB JAX directory) and Git LFS variable sizes (141.8 MB, 445.6 MB, 141.5 MB).

[1] [3] [4]

Is this for you?

Use cases
  • Running offline or closed-loop 3 Hz RGB + language manipulation inference with RT-1-X TensorFlow or JAX checkpoints
  • Benchmarking generalist cross-embodiment policies against a canonical Open X-Embodiment baseline
  • Inspecting or running the three original RT-1 TensorFlow SavedModels (700-task main, multi-robot EDR+Kuka, sim+real)

[1] [3]

Known limitations
  • google-research/robotics_transformer was archived on GitHub (last commit 2023-02-10, pushed 2024-01-31) and depends on legacy TensorFlow / TF-Agents / tensor2robot packages.
  • RT-1-X takes only a single workspace RGB camera and a task string; wrist cameras, in-hand cameras, depth, and proprioception are not supported as inputs.
  • Original RT-1 checkpoints in google-research/robotics_transformer are stored via Git LFS; RT-1-X checkpoints are hosted separately in a Google Cloud Storage bucket.

[1] [3] [4]

Safety and limitations
  • google-research/robotics_transformer is archived read-only and explicitly states it is not an officially supported Google product.
  • Single-camera RGB input only; lacks wrist-camera, depth, and proprioceptive feedback and runs at a low 3 Hz control rate.
  • Open X-Embodiment README licenses software under Apache-2.0 and non-software materials (including checkpoints/datasets) under CC-BY-4.0.

[1] [3]

Compute & fine-tuning

Inference compute
  • Documented hardware: Google Colab / workstation GPU (per Open X-Embodiment inference Colab)
  • Precision: fp32
  • Edge / consumer capable: Yes

Compact 0.78 GB SavedModel zip running at 3 Hz; exact minimum VRAM and latency_ms are not documented in README.

[3] [4] · Editorial assessment

Fine-tuning
  • Supported: No

Upstream repositories release architecture code and inference checkpoints/Colabs rather than a turnkey end-to-end RT-1-X fine-tuning script.

[1] [3]

Host computer

Compact transformer footprint (~0.78 GB RT-1-X TF SavedModel zip, ~1.37 GB JAX checkpoint, 0.14-0.45 GB RT-1 SavedModel variables) runnable on standard workstation GPUs or Colab.

Inferred from verified checkpoint sizes and Open X-Embodiment Colab inference notebook.

[1] [3] [4] · Editorial assessment

Operating system
  • Linux

TensorFlow/JAX/Bazel robotics tooling targets Linux.

[1] [3] · Editorial assessment

Dependencies
  • Python
  • TensorFlow & TF-Agents (for RT-1 / RT-1-X TF)
  • JAX & Flax (for RT-1-X JAX)
  • Bazel (optional, for running robotics_transformer tests)
  • Git LFS (for trained_checkpoints/ in robotics_transformer)

[1] [3]

Skills
  • Python
  • TensorFlow / SavedModel or JAX / Flax
  • gsutil / Google Cloud Storage downloads

Inferred from documented TensorFlow SavedModel, Bazel, and JAX/gsutil setup.

[1] [3] · Editorial assessment

Setup & usage

Open the upstream getting-started guide [3]

  • For RT-1-X TensorFlow inference, download rt_1_x_tf_trained_for_002272480_step.zip from GCS or follow the Open X-Embodiment inference Colab.
  • For RT-1-X JAX inference, download gs://gdm-robotics-open-x-embodiment/open_x_embodiment_and_rt_x_oss/rt_1_x_jax via gsutil and run models/rt1_inference_example.py from google-deepmind/open_x_embodiment.
  • For original RT-1 SavedModels, clone google-research/robotics_transformer with Git LFS, install requirements.txt, and load trained_checkpoints/{rt1main,rt1multirobot,rt1simreal} via tf.saved_model.load.

[1] [3] [4]

Upstream-reported benchmarks

Upstream-reported only: Open Source Robotics has not run training, inference, or hardware evaluations for these benchmarks.

BenchmarkUpstream-reported scoreScope & caveats
RT-1 (arXiv:2212.06817) and Open X-Embodiment RT-1-X (arXiv:2310.08864) multi-robot evaluationsSee arXiv:2212.06817 and arXiv:2310.08864 (no numeric table in repository READMEs)Upstream-reported multi-robot and cross-embodiment evaluations; not independently reproduced by OSR.

[5] [6]

Project notes

Architecture, observation space, and checkpoints

In transformer_network.py and configs/transformer_mixin.gin, RT-1 combines a FiLM-conditioned EfficientNet image tokenizer, TokenLearner compression, and an 8-layer decoder-only Transformer (num_layers: 8, num_heads: 8, layer_size: 128, token_embedding_size: 512, feed_forward_size: 512, vocab_size: 256, time_sequence_length: 6). Per the Open X-Embodiment README, RT-1-X takes a single workspace RGB camera and a language instruction—without wrist cameras, depth, or proprioception—under an assumed 3 Hz control cadence (333 ms update interval, not a measured inference latency). It outputs seven discretized gripper action variables (x, y, z, roll, pitch, yaw, and gripper opening) quantized into 256 bins; per the README, each variable may represent an absolute value, a delta change, or a velocity depending on the embodiment convention, which must be established before interpreting predictions. [1] [3]

Metadata checks confirmed five public checkpoint artifacts without downloading or loading model payloads: an HTTP HEAD 200 response for the RT-1-X TensorFlow SavedModel archive (rt_1_x_tf_trained_for_002272480_step.zip, 0.78 GB), a GCS object-listing byte total of 1.37 GB for the RT-1-X JAX/Flax directory, and Git LFS pointer metadata inside robotics_transformer/trained_checkpoints/ showing variable-file sizes of 0.14 GB (rt1main), 0.45 GB (rt1multirobot), and 0.14 GB (rt1simreal). [1] [3] [4]

Documented inference routes and upstream-reported evaluations

Because google-research/robotics_transformer is archived and neither repository provides a turnkey end-to-end RT-1-X fine-tuning pipeline, the documented workflows are inference examples—offline episode inference in the Open X-Embodiment TensorFlow Colab and the JAX/Flax checkpoint loader in models/rt1_inference_example.py—which also makes the checkpoints suitable as an editorial baseline reference. Exact scalar parameter counts, minimum GPU VRAM, and per-step latency numbers are not listed in either README, and multi-robot evaluation tables live in arXiv:2212.06817 and arXiv:2310.08864; Open Source Robotics has not run training, fine-tuning, inference, or robot evaluations for this profile. [1] [3] [5] [6]

What we checked

None. Prepared from public upstream sources without hands-on training, fine-tuning, or robot execution.

Updates

Upstream last updated
Jan 31, 2024 [1]
Latest release
Oct 10, 2023 [4]
Profile updated
Nov 21, 2026
Sources last checked
Nov 21, 2026
Next review
Feb 19, 2027
Reviewed revision
google-research/robotics_transformer@4569641b8111f3f402c32d8e24becd2a6e952ecc (archived); google-deepmind/open_x_embodiment@9eeb68b989efbcf474e8fb9019e01d02b962a604 · combined hardware and software repository

Separate license matrix

Foundational open-weights VLA model family with Apache-2.0 code and five ungated TensorFlow/JAX checkpoints across google-research/robotics_transformer and google-deepmind/open_x_embodiment.

LayerAvailability / license
Codeavailable · Apache-2.0
Verified Apache License Version 2.0 in google-research/robotics_transformer/LICENSE and google-deepmind/open_x_embodiment README. [1] [2] [3]
Model weightsavailable · Apache-2.0 (RT-1 checkpoints in robotics_transformer); CC-BY-4.0 (RT-1-X checkpoints under Open X-Embodiment non-software materials)
All five checkpoints (RT-1-X TF & JAX on GCS; rt1main, rt1multirobot, rt1simreal in Git LFS) are publicly downloadable without gating. [1] [2] [3] [4]
Training datasetsavailable · CC-BY-4.0 (Open X-Embodiment aggregate terms; individual constituent datasets carry their own licenses)
[3]
Documentationavailable · Apache-2.0 / CC-BY-4.0
[1] [3]

Licensing split across repositories

All software in both google-research/robotics_transformer and google-deepmind/open_x_embodiment is licensed under Apache-2.0. The Open X-Embodiment README further specifies that non-software materials (including the RT-1-X checkpoints and dataset distribution) are licensed under CC-BY-4.0, while individual constituent datasets within Open X-Embodiment carry their own upstream terms. [2] [3]

References

Each source links to the exact revision we read. “Used for” shows where this page relies on it.

Correction guidance · Discussion is not enabled · Profile format 0.8.0-draft.1

Change history

2026-11-21 — Initial 0.8.0-draft.1 model profile prepared from archived google-research/robotics_transformer@4569641b8111 and google-deepmind/open_x_embodiment@9eeb68b989ef.