octo Octo: An Open-Source Generalist Robot Policy

Octo Model Team
Dibya Ghosh*,1 Homer Walke*,1 Karl Pertsch*,1,2 Kevin Black*,1 Oier Mees*,1
Sudeep Dasari3 Joey Hejna2 Tobias Kreiman1 Charles Xu1 Jianlan Luo1 You Liang Tan1
Pannag Sanketi4 Quan Vuong4 Ted Xiao4 Dorsa Sadigh2 Chelsea Finn2 Sergey Levine1
*denotes equal contribution, listed in alphabetical order
1. UC Berkeley 2. Stanford University 3. Carnegie Mellon University 4. Google DeepMind

We introduce Octo octo, our ongoing effort for building open-source, widely applicable generalist policies for robotic manipulation. The Octo model is a transformer-based diffusion policy, pretrained on 800k robot episodes from the Open X-Embodiment dataset. It supports flexible task and observation definitions and can be quickly finetuned to new observation and action spaces. We are introducing two initial versions of Octo, Octo-Small (27M parameters) and Octo-Base (93M parameters).

teaser

The Model

The design of the Octo model emphasizes flexibility and scale: the model is designed to support a variety of commonly used robots, sensor configurations, and actions, while providing a generic and scalable recipe that can be trained on large amounts of data. Octo supports both natural language instructions and goal images, observation histories, and multi-modal action distributions via diffusion decoding. Furthermore, we designed Octo specifically to support efficient finetuning to new robot setups, including robots with different actions and different combinations of cameras and proprioceptive information. This design was selected specifically to make Octo a flexible and broadly applicable generalist robotic policy that can be utilized for a variety of downstream robotics applications and research projects. model

The Data

We train Octo on a mixture of 25 datasets from the Open X-Embodiment Dataset, a diverse collection of robot learning datasets. Our training mixture includes data from a variety of robot embodiments, scenes, and tasks. These datasets are heterogeneous not just in terms of the robot type, but also in the sensors (e.g., including or not including wrist cameras) and labels (e.g., including or not including language instructions). model

The Results

We evaluate Octo on 9 real robot setups across 4 institutions. Our evaluations capture diverse object interactions (e.g., "WidowX BridgeV2"), long task horizons (e.g., "Stanford Coffee") and precise manipulation (e.g., "Berkeley Peg Insert"). We evaluate Octo's capabilities to control robots in environments from the pretraining data out-of-the-box and to efficiently finetune to new tasks and environments with small target domain datasets. We also test finetuning with new observations (force-torque inputs for "Berkeley Peg Insert") and action spaces (joint position control in "Berkeley Pick-Up"). model

Zero-shot
WidowX UR5 RT-1 Robot
RT-1-X 0.20 0.35 0.60
RT-2-X 0.50 0.85
Octo octo 0.50 0.70 0.80
Finetuning
CMU Baking Stanford Coffee Berkeley Peg Insert* Berkeley Pick-Up Berkeley Bimanual Berkeley Coke Average
From Scratch 0.25 0.45 0.10 0.00 0.20 0.20 0.20
VC-1 0.30 0.00 0.05 0.00 0.50 0.10 0.15
Octo octo 0.50 0.75 0.70 0.60 0.80 1.00 0.72
*New observation input (force-torque proprioception)
New action space (joint position control)

Out-of-the-box, Octo can control multiple robots in environments from the pretraining data. When using natural language to specify tasks, it outperforms RT-1-X: the current best, openly available generalist robotic policy. It performs similarly to RT-2-X, a 55-billion parameter model. Additionally, while RT-1-X and RT-2-X only support language conditioning, Octo also supports goal image conditioning. On the WidowX tasks, we found that Octo achieved even better performance with goal image conditioning — 25% higher on average — likely because goal images provide more information about how to achieve the task.

We also find that finetuning Octo leads to better policies than starting from scratch or with the pretrained VC-1 weights. On average across the six evaluation setups, Octo outperforms the next best baseline by 52%. Each task uses ~100 target demonstrations. Importantly, we use the same finetuning recipe for all evaluation tasks, making this a good default configuration for Octo finetuning. The results also underline Octo’s ability to accommodate new observations (force-torque inputs for “Berkeley Insertion”), action spaces (joint position control for “Berkeley Pick-Up”) and new robot embodiments (“Berkeley Bi-Manual” and “Berkeley Coke”). This makes Octo applicable to a wide range of single and dual arm robotic manipulation problems that go beyond a single camera input and end-effector position control.

Citation

Please use the following BibTeX entry to cite this work:

@inproceedings{octo_2023,
    title={Octo: An Open-Source Generalist Robot Policy},
    author = {{Octo Model Team} and Dibya Ghosh and Homer Walke and Karl Pertsch and Kevin Black and Oier Mees and Sudeep Dasari and Joey Hejna and Charles Xu and Jianlan Luo and Tobias Kreiman and {You Liang} Tan and Pannag Sanketi and Quan Vuong and Ted Xiao and Dorsa Sadigh and Chelsea Finn and Sergey Levine},
    booktitle = {Proceedings of Robotics: Science and Systems},
    address  = {Delft, Netherlands},
    year = {2024},
}