Dyna-1 Base Model
AI model · Robotics / world model by Dyna Robotics
Zero-shot dexterity and one-hour fine-tuning for production-ready robots.
A foundation model for embodied AI that provides zero-shot dexterity for industrial tasks and reaches 100% success on new skills with just one hour of fine-tuning.
Model specs
The full spec card (context, pricing per token, modalities and more) on AntSpace. Sign up free to see the full spec card.
- Launched
- 2025-11-20
- Languages
- English
Category: Machine Learning
- Released
- 2025-11-20
- Fine-tuning Data Requirement
- 1 hour for ~100% success rate
- Zero-shot Tasks
- Laundry folding, package sorting
- Model Category
- Embodied AI Foundation Model
- Zero-shot dexterity for laundry folding and package sorting
- 100% success on new tasks with 1 hour of fine-tuning data
- Performs in unseen environments without task-specific tuning
- Foundation for production-ready embodied AI systems
Overview
The Dyna-1 Base Model is Dyna Robotics' foundation model designed for high-throughput dexterous autonomy in real-world environments. It serves as the core engine for robotic systems in hospitality, logistics, and factory operations, providing zero-shot capabilities for tasks like laundry folding and package sorting. The model is engineered to solve the problem of performance degradation in unseen environments, maintaining high reliability without environment-specific data. For businesses with unique requirements, it supports rapid customization, reaching near 100% success rates on new tasks with just one hour of demonstration data. This efficiency significantly reduces the commercial barrier to deploying advanced robotics for complex, long-horizon workflows.
Core Capabilities
- Environment Generalization: It achieves high performance on tasks even in environments it has never seen before, such as different room layouts or lighting conditions.
- Pre-training Autonomy: Behaviors are derived entirely from pre-training without the need for environment-specific data.
Fine-tuning and Customization
- 1-Hour Fine-tuning: The base model can be fine-tuned to reach roughly 100% success on challenging new tasks with as little as one hour of demonstration data.
- Task-Specific Training: The model has been successfully fine-tuned for complex activities such as cup stacking and celery chopping.
- Service Task Adaptation: It supports training for service-oriented tasks including breakfast making and tablemat preparation.
Use cases
- Logistics providers use the base model to sort packages in new warehouse layouts without needing site-specific training data.
- Hospitality businesses deploy the model for laundry folding tasks that adapt to different towel types and folding standards.
- Food service operators fine-tune the model in one hour to handle specialized tasks like celery chopping or meal preparation.
Specifications
- Fine-tuning
- Supports extremely efficient fine-tuning with as little as one hour of demonstration data
- Released
- 2025-11-20
- Fine-tuning Data Requirement
- 1 hour for ~100% success rate
- Zero-shot Tasks
- Laundry folding, package sorting
- Tested Skills
- Cup stacking, celery chopping, breakfast making, tablemat preparation
- Model Category
- Embodied AI Foundation Model
How it works
- General Pre-training: The model undergoes extensive pre-training to develop general dexterous capabilities that work out-of-the-box for common tasks.
- Data Collection: For specialized or new tasks, users provide approximately one hour of demonstration data to the base model.
- Efficient Fine-tuning: The model uses the demonstration data to reach high reliability (roughly 100% success) for the specific commercial requirement.
FAQ
How much data is needed to fine-tune the model for a new task?
The base model can reach roughly 100% success on new tasks with as little as one hour of demonstration data.
Does the model work in unseen environments?
Yes, the model can perform diverse dexterous tasks in environments it has never seen before without any task-specific fine-tuning.
What tasks has the Dyna-1 Base Model been tested on?
The model has been stress-tested on tasks including laundry folding, package sorting, cup stacking, celery chopping, breakfast making, and tablemat preparation.
Added by AntSpace from public sources (dyna.co).