Octo
Octo Model TeamOcto is an open generalist robot policy from UC Berkeley's Robotic AI and Learning Lab and collaborators: a transformer diffusion policy trained on 800k trajectories from Open X-Embodiment. It accepts several camera views and either a language instruction or a goal image, and its modular attention lets it be fine-tuned to new sensors, action spaces and robot bodies with a small dataset.
Development stopped after the 1.5 release in mid-2024.
Openness
5 high confidence- weights
- open
- data
- open(the named Open X-Embodiment mixture)
- code
- open(pretraining and fine-tuning)
- license
- MIT(code and checkpoints)
Everything needed to rebuild Octo is public: MIT weights and code, the pretraining and fine-tuning scripts, and a card that lists each Open X-Embodiment dataset in the training mix with its share.
- https://cdn.jsdelivr.net/gh/octo-models/octo/LICENSE recorded 2026-09-26
LICENSE: MIT License, "Copyright (c) 2023 Robotic AI & Learning Lab Berkeley".
- https://huggingface.co/api/models/rail-berkeley/octo-base-1.5 recorded 2026-09-26
Hub API for rail-berkeley/octo-base-1.5: license:mit, gated false.
- https://huggingface.co/rail-berkeley/octo-base-1.5/raw/main/README.md recorded 2026-09-26
octo-base-1.5 card: "trained on a mix of datasets from the Open X-Embodiment dataset" with the mixture listed by weight (Fractal, Kuka, Bridge 17% each, BC-Z 9.1%, and more).
- https://ungh.cc/repos/octo-models/octo/files/main recorded 2026-09-26
Repository tree: scripts/train.py, scripts/finetune.py, scripts/configs/octo_pretrain_config.py and octo/data/oxe/oxe_dataset_mixes.py, which defines the Open X-Embodiment training mixes.
Adoption
1 medium confidenceHugging Face downloads summed over the four Octo checkpoints.
- https://huggingface.co/api/models?author=rail-berkeley&sort=downloads&limit=50&expand[]=downloads&expand[]=likes&expand[]=cardData&expand[]=lastModified&expand[]=gated recorded 2026-09-26
Hub listing for rail-berkeley: octo-small-1.5 346, octo-small 153, octo-base-1.5 128, octo-base 116 downloads in 30 days.
Capability
3 medium confidenceOcto, like OpenVLA, learned from pooled data across many single-arm robots, and it was built to be fine-tuned onto new bodies and sensors. At under a hundred million parameters it has no language-model backbone, and it ships no bimanual or humanoid policy.
- https://cdn.jsdelivr.net/gh/octo-models/octo/README.md recorded 2026-09-26
README: "transformer-based diffusion policies, trained on a diverse mix of 800k robot trajectories"; "can control various robot arms"; "effectively finetuned to robot setups with new sensory inputs, action spaces, and morphologies".
- https://huggingface.co/rail-berkeley/octo-base-1.5/raw/main/README.md recorded 2026-09-26
Card: "The model is a Transformer with 93M parameters (equivalent to a ViT-B)".
Verified 2026-09-26