Octo: An Open-Source Generalist Robot Policy

Octo addresses a critical gap in robot foundation models: while proprietary generalist policies like RT-2 showed impressive results, there was no open-source alternative that researchers could freely fine-tune, inspect, and modify. Octo fills this niche. Trained on 800,000 trajectories from the Open X-Embodiment dataset, Octo is a transformer-based policy capable of conditioning on language commands or goal images and generating actions for diverse sensor and actuator configurations. It supports both visual and proprioceptive observations and can output continuous robot actions at interactive rates. The architecture is designed for practical flexibility. Unlike prior work that couples the policy tightly to specific robot hardware, Octo employs an observation-action tokenizer that projects heterogeneous sensor streams into a common token space. This means the same pretrained backbone can be fine-tuned to a new robot with different camera counts, image resolutions, or action dimensions in just a few hours on a single consumer GPU. The authors demonstrate this cross-platform capability on 9 robotic platforms, ranging from low-cost WidowX arms to Franka Research robots and mobile manipulators, confirming that Octo serves as an effective policy initialization across the board. Training at this scale required careful engineering. The authors release their full pre-training pipeline, including data loading, tokenization, and distributed training scripts, all built on top of JAX and Flax. They conduct extensive ablations on architectural choices—such as whether to condition on language tokens early or late in the transformer stack—and on dataset curation, finding that data diversity and quality filtering matter more than sheer scale for generalization. These ablation studies provide valuable guidance for future efforts in building generalist robot models. Empirically, Octo achieves competitive or superior performance compared to prior open methods and serves as a strong initialization for downstream tasks. When fine-tuned on a small amount of in-domain data, it quickly adapts to new manipulation skills, object geometries, and environments that were not present in the pre-training dataset. The authors also provide a Hugging Face integration, making it possible to download pretrained weights and fine-tune models with a few lines of Python. For the robot learning community, Octo represents a foundational open generalist policy that can be used as a starting point for a wide variety of research projects.