Return list
ENGINEAI Goes Open-Source: Sharing the Power of Innovation with the World!
Media Coverage2025-06-10
Deployment Code Usage Guide
训练框架
Training Framework
Designed as a versatile and general reinforcement learning (RL) framework for legged robots developed by ENGINEAI, the EngineAI RL Workspace integrates end-to-end modules for environment setup, algorithm training, and performance evaluation. With its modular design, highly scalable architecture, and full lifecycle toolchain, it provides developers with a seamless workflow from algorithm development to real-world validation, accelerating the iteration and deployment of reinforcement learning algorithms for legged robots.
1.Core Features
•
Environment Ecosystem: Includes modular components such as environment observation, domain randomization, goals, and reward functions, enabling rapid construction of diverse training scenarios； •
Algorithm Engine: Integrates algorithm executors, RL algorithms, neural network architectures, and efficient data storage components, allowing flexible algorithm modifications； •
Toolkit: A collection of utilities and reusable classes that avoid circular library dependencies； •
Module Integration: Calls functions from other modules at the integration layer to implement experiment registration, training, and evaluation.
•
Unified Algorithm Executor: Training and inference share the same execution logic — developers avoid rewriting base code and focus directly on algorithmic innovation； •
Algorithm-Environment Decoupling: Algorithm iterations require no changes to environment interfaces, only adjustments to input/output handling — drastically reducing development costs.
•
Dynamic Recording: Supports video capture of training/inference processes for real-time tracking of robot performance； •
Variable Control: Allows custom environment parameters and data logging, enabling controlled experiments through single-variable adjustments； •
Version Snapshots: Automatically saves code and configuration files during training, creating traceable experiment records.
•
Smart Versioning: Automatically loads code versions matching the experiment when resuming training, eliminating manual searches and ensuring reproducibility； •
Git Integration: Automatically records Git commit information for full code change traceability.
•
JSON One-Click Generation: Experiment parameters can be saved as .json files and converted to .py scripts via built-in tools, enabling instant task setup from historical templates and streamlined parameter tuning.
•
Cross-Platform Deployment: Provides tools to convert trained .pt models to ONNX, MNN, and other deployment formats, accelerating the path from training to simulation and real-robot deployment.
•
High-Level Development: Simplified control based on existing code — issuing body velocity commands (forward/backward speed, turning rate, etc.) activates basic robot mobility； •
Low-Level Development: Full customization via joint angle/torque commands, compatible with RL, traditional control, and other algorithms for precise motion control in complex scenarios.
1.
Empower Developers: drastically reduce robotics R&D costs and barriers, enabling startups, research institutions, and individual developers to accelerate personalized solutions； 2.
Ignite Industry Innovation: gather global expertise to breakthrough technical bottlenecks in key areas like education, commercial services, and industrial manufacturing； 3.
Accelerate Cross-Sector Collaboration: provide standardized, modular technical support to help industries achieve intelligent transformation and seamless robotics integration.
