Open Source RoboticsRobots, software & learning

Octo

Robotics model

Transformer diffusion robot policies with pretrained checkpoints and fine-tuning tools; observation history, action statistics and robot-specific interfaces must match the selected model. [1]

Setup & usage →CheckpointsModel hub / homepageGetting started

For
Robot-learning researchers with a compatible environment and accelerator setup [1]
Parameters
93M (Octo-Base-1.5) / 27M (Octo-Small-1.5) per model cards [8] [9]
Embodiments
WidowX (BridgeData V2), ALOHA (via fine-tuning example), Open X-Embodiment 25-dataset mixture robots (Franka, UR5,… [7] [8]
Source rights
Some licenses are unresolved. Review licenses

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

Licenses & related hardware
License
MIT · license not fully assessed for datasets, documentation · Details by part

Requirements · Research details and dates

On this page

Overview

Octo (octo-models/octo, release v1.5) is a family of generalist transformer diffusion visuomotor policies trained on a 25-dataset mixture from Open X-Embodiment. Two ungated JAX Orbax checkpoints are published on Hugging Face Hub: rail-berkeley/octo-base-1.5 (93M parameters, ViT-B equivalent, 0.75 GB on disk) and rail-berkeley/octo-small-1.5 (27M parameters, ViT-S equivalent, 0.51 GB on disk). Both tokenize images via a lightweight CNN into 16x16 patches, encode language with T5-Base, and predict continuous actions with a diffusion head. [1] [2] [8] [9] [10]

Model card summary

Architecture

Transformer-based generalist diffusion policy (ViT-B 93M for Base, ViT-S 27M for Small) using a lightweight CNN image tokenizer (16x16 patches), T5-Base language encoder, and diffusion action head.

[8] [9]

Base model

T5-Base language encoder (vision encoder and transformer trained from scratch on OXE)

[8] [9]

Parameter count

93M (Octo-Base-1.5) / 27M (Octo-Small-1.5) per model cards

[8] [9]

Input modalities
rgblanguage

Primary RGB (256x256), wrist RGB (128x128), goal images, and language instructions (history window up to 2 timesteps); fine-tuning can add proprioception.

[8] [9]

Output modalities
ee-pose-deltagripper-command

4-step chunks of 7-dimensional actions.

[1] [8]

Action space

Continuous diffusion action chunks predicting 4 future steps of 7-dimensional actions (window size 2; translation, rotation, and gripper in the Bridge example), unnormalized via dataset/robot-specific statistics; supports full-chunk, first-step, or temporal-ensembling execution.

[1] [5] [8]

Trained / supported embodiments
  • WidowX (BridgeData V2)
  • ALOHA (via fine-tuning example)
  • Open X-Embodiment 25-dataset mixture robots (Franka, UR5, Kuka, Google Robot, Stretch, Jaco, etc.)

[7] [8]

Training datasets
  • Open X-Embodiment 25-dataset mixture (Fractal, Kuka, Bridge, BC-Z, Stanford Hydra, Language Table, etc.)

[8]

Interfaces
  • OctoModel.load_pretrained
  • create_tasks for language or goal-image task input
  • sample_actions for action chunks
  • Gym environment adapters

[1] [5] [6]

Checkpoints

CheckpointFormatSizeAccessPinned revision
rail-berkeley/octo-base-1.5jax-orbax0.75 GBopenee3c10e8edd6
rail-berkeley/octo-small-1.5jax-orbax0.51 GBopendc9aa3019f76

Ungated JAX Orbax checkpoints verified via Hugging Face API.

[8] [9] [10]

Is this for you?

Use cases
  • Zero-shot language- or goal-image-conditioned manipulation on OXE-matched setups
  • Fine-tuning compact diffusion policies for new observation/action spaces (e.g., ALOHA)

[1] [7]

Known limitations
  • Action chunks are not complete task trajectories
  • Missing modalities and absent history require appropriate masks
  • Pretraining dataset and compute are substantially larger than the inference example
  • Model and source licenses do not establish a uniform license for every training dataset

Last item is an editorial rights-scope limitation.

[1] [5] [8]

Safety and limitations
  • 4-step action chunks are not complete task trajectories; execution requires matching history masks (timestep_pad_mask) and per-modality masks.
  • Pinned to Python 3.10 and JAX 0.4.20 (CUDA 11 or TPU), which requires care on newer CUDA/driver stacks.
  • MIT code and checkpoint licenses do not establish a uniform license across the 25 constituent Open X-Embodiment datasets.

[1] [3] [5] [8]

Compute & fine-tuning

Inference compute
  • Documented hardware: CUDA 11 GPU or TPU (JAX 0.4.20; 0.51-0.75 GB checkpoints)
  • Precision: fp32
  • Edge / consumer capable: Yes

Compact 27M-93M parameter footprint (0.51-0.75 GB on disk); edge/single-GPU inference capability is editorially inferred while exact minimum VRAM and latency_ms are not stated.

[1] [10] · Editorial assessment

Fine-tuning

Debug fine-tuning example loads octo-small-1.5; ALOHA example adds separate simulator and ACT dependencies.

[1] [7]

Host computer

CUDA 11 GPU or TPU with JAX 0.4.20 for inference and fine-tuning; upstream pretraining used ~1.2 TB prepared data on TPU pods.

[1]

Operating system
  • Linux

CUDA 11 / TPU JAX Linux setup documented.

[1] · Editorial assessment

Dependencies
  • Python 3.10 in upstream environment instructions
  • JAX 0.4.20 with CUDA11 or TPU installation route
  • Checkpoint and robot/dataset-specific normalization statistics
  • Separate robot or Gym-compatible evaluation environment

[1] [5] [6]

Skills
  • Python
  • JAX
  • Gym environment wrappers

Inferred from documented Python/JAX and Gym adapter instructions.

[1] [6] · Editorial assessment

Setup & usage

Open the upstream getting-started guide [5] · Model card / documentation [1] [4]

  • Create Python 3.10 environment and install JAX 0.4.20 (CUDA 11 or TPU route)
  • Load a pinned v1.5 checkpoint (octo-base-1.5 or octo-small-1.5) via OctoModel.load_pretrained
  • Validate observation keys, history masks (timestep_pad_mask), and dataset unnormalization statistics using 01_inference_pretrained.ipynb

[1] [5]

Upstream-reported benchmarks

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

BenchmarkUpstream-reported scoreScope & caveats
Octo upstream paper and repository evaluationsUpstream reports multi-embodiment zero-shot and fine-tuning results in the project paper and README (specific scores not reproduced here)Upstream-reported evaluation claims; not independently verified by OSR.

[1]

Project notes

Observation history, masks, and action chunks

Both v1.5 checkpoints accept primary RGB images (256x256), wrist RGB images (128x128), goal images, and language instructions over a history window of up to 2 timesteps, predicting 4-step chunks of 7-dimensional actions. In the supplied Bridge inference notebook (01_inference_pretrained.ipynb), those seven components map to end-effector translation, rotation, and gripper control; target environments must match the selected checkpoint's action convention and dataset unnormalization statistics rather than assuming identical semantics across robots. At inference, any subset of observation and task keys may be passed, but timestep_pad_mask must mark absent history steps separately from missing camera modalities, and chunks can be executed via full-chunk, first-action, or temporal-ensembling loops. Consult the Gym adapter guide (examples/envs/README.md) as the next integration reference when wiring a new environment. [1] [5] [6] [8]

Environment, fine-tuning, and upstream-reported benchmarks

Upstream setup specifies Python 3.10 and JAX 0.4.20 with CUDA 11 or TPU routes. Upstream documents GPU/TPU inference as well as fine-tuning workflows for new observation and action spaces (including an ALOHA example in 03_eval_finetuned.py with separate simulator and ACT dependencies), while full pretraining from scratch uses ~1.2 TB of prepared data on TPU pods. Minimum VRAM, inference latency, and target-robot execution timing/control frequency are not established here—checkpoint download sizes (0.51–0.75 GB) are disk sizes rather than runtime memory requirements, and single-GPU suitability for fine-tuning is not established by the inspected sources. Upstream evaluation claims in the project paper and README are upstream-reported; Open Source Robotics has not downloaded checkpoints or run training, fine-tuning, inference, or robot evaluations. [1] [7] [10]

What we checked

Source research only; no installation, physical build or performance test performed. No affiliate links added.

Not yet established

  • Documented closed-loop control frequency (Hz)

Updates

Upstream last updated
May 23, 2024 [2]
Latest release
May 24, 2024 [2]
Profile updated
Nov 21, 2026
Sources last checked
Nov 21, 2026
Next review
Feb 19, 2027
Reviewed revision
5eaa5c6960398925ae6f52ed072d843d2f3ecb0b (release v1.5) · software repository

Separate license matrix

Available project sources support research coverage; complete build and commercial qualification remain separate.

LayerAvailability / license
Codeavailable · MIT
[3]
Model weightsavailable · MIT
Pinned Octo Base 1.5 (ee3c10e8) and Octo Small 1.5 (dc9aa301) model cards both declare MIT; checkpoint payloads not downloaded or executed. [1] [8] [9] [10]
Training datasetsavailable · License not fully assessed
Upstream links the Open X-Embodiment collection and preparation script. Constituent dataset licenses and payloads were not individually verified; no uniform dataset license asserted. [1] [8]
Documentationavailable · License not fully assessed
Examples and source documentation available; separate documentation-wide license not established. [1] [4] [5]

Separate code, weight, and dataset rights

The octo-models/octo repository and both pinned v1.5 model cards (octo-base-1.5 and octo-small-1.5) declare MIT licenses. However, the 25 constituent Open X-Embodiment datasets retain their own upstream licenses, which were not individually verified here. [1] [3] [8] [9]

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 — Migrated 0.6.0-draft.1 -> 0.7.0-draft.5 -> 0.8.0-draft.1 (kind=model); added verified Small 1.5 card (S009) and HF API metadata (S010) while preserving historical source dates.