Octo Base

About UC Berkeley / Stanford (Octo Model Team)

Founded 2023 · Berkeley & Stanford, California, USA

The Octo project is a collaboration of academic labs led by Sergey Levine (UC Berkeley BAIR) and Chelsea Finn (Stanford IRIS), with contributions from CMU, Google DeepMind, and Toyota Research Institute. Octo was first released in May 2024 alongside the Open-X-Embodiment dataset effort, with the goal of producing a generalist, fully open-source robot policy that any researcher can fine-tune on a new robot in hours. Octo introduced the recipe of a transformer policy with a diffusion action head trained on 800k cross-embodiment demonstrations, and it has become a de-facto baseline in academic VLA / generalist-policy research. The team released both Octo-Small (27M) and Octo-Base (93M) under Apache-2.0, alongside code, checkpoints and a fine-tuning toolkit.

Architecture

Transformer policy with diffusion action head (Vision-Language-Action)

Octo-Base is a transformer-based generalist robot policy. Inputs are tokenised RGB views and a natural-language instruction (encoded with a T5-base text encoder), interleaved with learnable readout tokens. The transformer trunk consumes this sequence and emits action latents that are decoded by a diffusion head producing continuous action chunks (default 4-step lookahead, 7-DoF end-effector deltas). The model was pretrained on roughly 800k demonstrations from 25 datasets in the Open-X-Embodiment collection, covering 9 robots, both single-arm and bimanual setups. Octo is intentionally embodiment-agnostic: action and proprioception spaces are encoded via shared adapters so the same backbone can be fine-tuned to new robots with as little as a few hundred demos. The diffusion head gives smooth, multimodal trajectories that outperform discrete-token VLAs on dexterous tasks at this scale.

Parameters

93M

Capabilities

• Generalist VLA policy across many robot embodiments
• Trained on ~800k demos from Open-X-Embodiment
• Diffusion action head produces smooth continuous actions
• Natural-language instruction conditioning (T5 encoder)
• Multi-view image inputs (primary + wrist cameras)
• Designed for fast fine-tuning on new robots and tasks
• Apache-2.0 open weights, code and recipes
• Strong academic baseline for VLA papers
• Best for: research, fine-tuning to new embodiments, generalist-policy benchmarks.

Training & license

~800,000 robot trajectories drawn from 25 Open-X-Embodiment-compatible datasets across 9 robot embodiments (Franka, WidowX, Bridge, RT-1 Everyday Robots, Berkeley UR5, etc.). Trained on TPU v4 / v5 hardware.

License: Apache-2.0 - fully open weights, code, and dataset references. Research-friendly; commercial use permitted under the licence.

Safety testing: Octo is a research artifact; no formal red-teaming or RSP. Risks (unsafe robot motions) are mitigated at deployment time by application-specific safety controllers, not by the model itself.