Octo Small

Um UC Berkeley / Stanford (Octo Model Team)

Stofnað 2023 · Berkeley & Stanford, California, USA

Octo-Small is the compact 27M-parameter variant of the Octo generalist robot policy released by the Octo Model Team - a UC Berkeley + Stanford-led collaboration (Sergey Levine and Chelsea Finn labs) with contributors from Google DeepMind, CMU and Toyota Research Institute. Octo-Small was released alongside Octo-Base in May 2024 to give researchers a CPU/edge-friendly option that still benefits from the same 800k Open-X-Embodiment pretraining recipe. It is widely used in academic teaching, robotics coursework, and rapid prototyping where the 93M Base model is too heavy for the available hardware.

Arkitektúr

Compact transformer policy with diffusion action head (Vision-Language-Action)

Octo-Small is architecturally identical to Octo-Base but uses a smaller transformer trunk (~27M parameters total). Inputs are tokenised RGB observations from primary and wrist cameras and a T5-base-encoded language instruction, plus learnable readout tokens. The transformer fuses these tokens and emits action latents that are decoded by a diffusion head into continuous 7-DoF end-effector action chunks. It is pretrained on the same ~800k demonstrations from 25 Open-X-Embodiment-compatible datasets across 9 robot embodiments. Despite being ~3.4x smaller than Octo-Base, the small variant retains the diffusion-policy output and the embodiment-agnostic input adapters, making it suitable as a fast baseline and a starting point for fine-tuning to new robots on modest GPUs.

Breytur

27M

Geta

• Compact 27M generalist VLA policy
• Same training recipe and dataset as Octo-Base
• Continuous action chunks via diffusion head
• Runs on a single consumer GPU and many edge devices
• Fast fine-tuning on new robots / tasks
• Natural-language instruction conditioning
• Apache-2.0 open weights and code
• Reproducible baseline for academic VLA work
• Best for: edge robotics, teaching, fast iteration.

Þjálfun og leyfi

~800,000 cross-embodiment robot demonstrations from 25 Open-X-Embodiment datasets (same corpus as Octo-Base). Trained on TPU hardware with the public Octo recipe.

Leyfi: Apache-2.0 - fully open weights, code, and recipes. Research and commercial use permitted under the licence.

Öryggispróf: Research artifact; no formal RSP. Operational safety is the responsibility of downstream deployers (e.g. safety controllers, force limits).