AgentSkillsCN

Reinforcement Learning Skill

通过仿真训练，结合从仿真到现实的迁移策略，开展机器人控制的强化学习训练

SKILL.md
--- frontmatter
name: Reinforcement Learning Skill
description: RL training for robot control using simulation with sim-to-real transfer
slug: rl-robotics
category: Learning
allowed-tools:
  - Bash
  - Read
  - Write
  - Edit
  - Glob
  - Grep

Reinforcement Learning Skill

Overview

Expert skill for training reinforcement learning agents for robot control tasks, including environment design, training pipelines, and sim-to-real transfer.

Capabilities

  • •Configure Gym/Gymnasium environments for robots
  • •Set up Stable Baselines3 training (PPO, SAC, TD3)
  • •Implement custom observation and action spaces
  • •Design reward shaping strategies
  • •Configure parallel environment training
  • •Implement domain randomization for sim-to-real
  • •Set up curriculum learning
  • •Configure vision-based RL with CNNs
  • •Implement policy distillation
  • •Export policies for deployment (ONNX, TorchScript)

Target Processes

  • •rl-robot-control.js
  • •imitation-learning.js
  • •sim-to-real-validation.js
  • •nn-model-optimization.js

Dependencies

  • •Stable Baselines3
  • •Gymnasium
  • •Isaac Gym
  • •rsl_rl

Usage Context

This skill is invoked when processes require RL-based robot control, learning from simulation, or transferring learned policies to real robots.

Output Artifacts

  • •Gymnasium environment implementations
  • •Training configurations
  • •Reward function designs
  • •Domain randomization configs
  • •Trained policy checkpoints
  • •Deployment-ready models (ONNX)