| Property | Value |
|---|---|
| Parameter Count | 193M |
| License | Apache 2.0 |
| Paper | arXiv:2402.09844 |
| Architecture | Transformer-based Multi-task Agent |
What is JAT?
JAT (Jack of All Trades) is a versatile transformer-based model designed for multi-task reinforcement learning. It represents a significant advancement in creating general-purpose AI agents capable of handling diverse tasks across multiple environments including Atari games, BabyAI navigation tasks, MetaWorld manipulation tasks, and MuJoCo physics simulations.
Implementation Details
The model utilizes a transformer architecture with 193M parameters, trained on a comprehensive suite of tasks. It demonstrates remarkable adaptability across different domains while maintaining F32 precision for optimal performance.
- Achieves 0.14 IQM expert normalized reward on Atari-57 benchmark
- Reaches 0.99 IQM expert normalized reward on BabyAI tasks
- Attains 0.65 IQM expert normalized reward on MetaWorld challenges
- Performs at 0.85 IQM expert normalized reward on MuJoCo environments
Core Capabilities
- Multi-environment mastery across 57 Atari games
- Advanced navigation and instruction following in BabyAI
- Robotic manipulation tasks in MetaWorld
- Complex physics-based control in MuJoCo
Frequently Asked Questions
Q: What makes this model unique?
JAT stands out for its ability to handle multiple types of reinforcement learning tasks within a single model architecture, demonstrating strong performance across diverse environments without task-specific adjustments.
Q: What are the recommended use cases?
The model is particularly suited for research in multi-task reinforcement learning, robotics simulation, game playing, and general AI agent development. It can serve as a foundation for developing more specialized agents or studying transfer learning across different domains.