Description
Octo represents an ongoing initiative to develop open-source, widely applicable generalist policies for robotic manipulation. The core of Octo is a transformer-based diffusion policy, meticulously pretrained on 800,000 robot episodes sourced from the comprehensive Open X-Embodiment dataset. This extensive pretraining equips Octo with a robust foundation for understanding and executing a diverse range of robotic tasks.
Read More
The design philosophy behind Octo emphasizes both flexibility and scalability. It is engineered to accommodate a variety of commonly utilized robots, sensor configurations, and action spaces. This generic and scalable architecture allows for training on massive datasets, leading to a highly adaptable model. Octo supports multiple input modalities for task specification, including natural language instructions and goal images. It can also process observation histories and generate multi-modal action distributions through diffusion decoding.
A key strength of Octo lies in its support for efficient finetuning. This capability allows the policy to be quickly adapted to new robot setups, including those with different action spaces or varied combinations of cameras and proprioceptive sensors. This design choice makes Octo a versatile and broadly applicable generalist robotic policy, suitable for numerous downstream robotics applications and research projects.
Octo is trained on a diverse mixture of 25 datasets from the Open X-Embodiment Dataset, encompassing a wide array of robot embodiments, scenes, and tasks. The heterogeneity of these datasets, in terms of robot types, sensors (e.g., with or without wrist cameras), and labels (e.g., with or without language instructions), contributes to Octo's robust generalization capabilities.
Evaluations of Octo have been conducted across nine real-world robot setups at four different institutions. These evaluations cover diverse object interactions, long task horizons, and precise manipulation tasks. Octo demonstrates strong out-of-the-box performance in environments from its pretraining data and excels in efficient finetuning to new tasks and environments with small target domain datasets. It also shows remarkable adaptability to new observations, such as force-torque inputs, and new action spaces, like joint position control, making it a powerful tool for advanced robotics research and development.
- CategoryRobotics & AI Hardware
- Pricing ModelOpen source
- DeploymentCloud
- PlatformsWeb
Octo Robot Policy Highlights
Transformer-based diffusion policy architecture
Pretrained on 800,000 robot episodes from Open X-Embodiment dataset
Supports natural language instructions for task specification
Supports goal image conditioning for task specification
Accommodates observation histories
Generates multi-modal action distributions via diffusion decoding
Designed for efficient finetuning to new observation and action spaces
Available in two versions: Octo-Small (27M parameters) and Octo-Base (93M parameters)
Evaluated on 9 real robot setups across 4 institutions
Outperforms RT-1-X in zero-shot language conditioning
Comparable performance to RT-2-X
Achieves higher performance with goal image conditioning on WidowX tasks
Demonstrates superior finetuning performance compared to training from scratch or with VC-1 weights
Getting Started with Octo Robot Policy
Access Model: Obtain the Octo model weights and code from the provided repositories.
Set Up Environment: Configure your robotics simulation or hardware environment.
Integrate Policy: Load the Octo model and define your observation and action spaces.
Task Specification: Provide task instructions via natural language or goal images.
Robot Control: Execute the policy to control robot actions.
Finetune Model: Adapt Octo to new tasks or environments using target demonstrations.
Octo Robot Policy's Use Cases
- General Robotic Manipulation
- Task-Specific Finetuning
- Multi-Modal Task Instruction
- Robotics Research Platform
- Sensor Configuration Adaptation
- Action Space Flexibility
