DEPLOYDatabase

Brain

Octo

Open-source generalist robot policy developed at UC Berkeley/Stanford. Transformer-based diffusion policy trained on 800K trajectories from the Open X-Embodiment dataset. Supports multi-robot and language-conditioned manipulation tasks. 27M and 93M parameter variants.

Research model · Maturity: Research · Open source

Machine-readable surfaces

Architecture

Transformer diffusion policy, 27M/93M params

Key facts

Octo model and data (primary project page)
The Octo project page describes Octo as a transformer-based diffusion policy pretrained on 800k robot episodes from the Open X-Embodiment dataset, trained on a mixture of 25 datasets. It introduces Octo-Small (27M parameters) and Octo-Base (93M parameters). Authors are affiliated with UC Berkeley, Stanford, Carnegie Mellon, and Google DeepMind. The citation is Robotics: Science and Systems 2024, Delft (Octo project page)
Reported zero-shot and finetune results (primary project page)
The same page reports zero-shot success of 0.50 on WidowX, 0.70 on UR5, and 0.80 on the RT-1 robot, and an average finetune success of 0.72 across six setups that each use about 100 target demonstrations, versus 0.20 from scratch. Treat these as paper-reported evaluation rates, not a deployed fleet census (Octo project page)
Open source code and checkpoints
The Octo project links an open source repository with model code, pretrained checkpoints, and fine tuning materials for the Octo generalist robot policy (Octo repository)
Flexible task and observation interface
The Octo project describes a policy designed to accept flexible task and observation definitions and adapt to new robotic platforms through fine tuning (Octo project)
Generalist policy architecture
The Octo team describes Octo as a transformer-based diffusion policy for broadly applicable robotic manipulation (Octo).
Pretraining corpus
The Octo team says the model was pretrained on 800,000 robot episodes from the Open X-Embodiment dataset (Octo).
Multimodal conditioning
The Octo team says Octo supports natural-language instructions, goal images, observation histories, and multimodal action distributions (Octo).
Fine-tuning flexibility
The Octo team says the policy can be fine-tuned to new observation and action spaces and different robot setups (Octo).
Published model sizes
The Octo project identifies Octo-Small at 27 million parameters and Octo-Base at 93 million parameters (Octo).

Mentioned on (1)

Registry pages whose text names Octo.

AI systems (1)

OpenVLA

Common questions

What is Octo?
Open-source generalist robot policy developed at UC Berkeley/Stanford. Transformer-based diffusion policy trained on 800K trajectories from the Open X-Embodiment dataset. Supports multi-robot and language-conditioned manipulation tasks. 27M and 93M parameter variants.
Is Octo open source?
Yes. Octo is recorded as open-source / open-weights on the DEPLOY registry, meaning model weights or source code are publicly available.
What type of AI is Octo?
Octo is a research model, built on a Transformer diffusion policy, 27M/93M params architecture on the DEPLOY registry.
What is Octo's maturity stage?
Octo is at the research stage on the DEPLOY maturity ladder. Research stage means active development without commercial deployments on file.
Which robots run on Octo?
No robot models on the DEPLOY registry are recorded as running Octo. DEPLOY wires brain-to-model connections only when the wiring is verifiable from primary sources; absence may reflect pre-deployment or unverified manufacturer claims.

Sources (4)

  1. https://octo-models.github.io/
  2. https://github.com/octo-models/octo
  3. https://octo-models.github.io/paper.pdf
  4. https://arxiv.org/html/2405.12213v2
Methodology: Verified · 4 sources (no primary) · last reviewed 2026-10-09

Verification posture

Verified

Low confidence

Review state

Stable

Last reviewed 2026-10-09

Maturity + lifecycle

Maturity stage: research

Sources by quality tier

2
unclassified
Unclassified source
1
code-repository
Code repository
1
preprint
Preprint

The framework is documented at /methodology. Corrections at /corrections. Reviewer: DEPLOY editorial team.

Methodology surface for Octo.

Canonical ID fe8e8e84-03cc-4745-8644-5729890a4103