drlzh.ai

What It Is

Deep Reinforcement Learning: Zero to Hero is a hands-on course built around deep reinforcement learning. The page highlights a “full pipeline” approach: learners work through 18 notebooks, filling in # TODO code blocks to implement algorithms such as DQN, SAC, PPO, AlphaZero, Dreamer, and GRPO from scratch, then train agents in scenarios including Atari, lunar landing, robotics, and LLM alignment. It feels more like an open-source course and coding lab than a traditional video course.

Core Content and Format

The course is divided into four stages: Foundations, Deep RL Core, Advanced, and Frontier. Topics start with MDPs, Gym, DQN, Policy Gradient, Actor-Critic, and PPO, then expand into exploration, multi-agent RL, offline/imitation learning, MCTS/AlphaZero, and further into RLHF, Decision Transformers, VLA, Productionizing RL, Dreamer, and Meta-RL. The main learning format is self-paced notebook study plus code-completion practice, with a solutions folder included. The page does not show any live classes, recorded lectures, or 1-on-1 services.

AI Companion and Learning Experience

A standout feature is DRL-ZH Companion, a VS Code extension. It observes the current TODO task, detects states such as idle, stuck, reading, confusion, drift, and flow, and provides Socratic-style hints without directly revealing the answer. Voice mode uses Whisper STT and local Kokoro TTS. For LLM support, users must provide their own Groq, Gemini, OpenAI, or Anthropic key, and cover the related costs themselves. The environment can be launched with Docker and three git commands, making the engineering setup relatively friendly.

Pricing, Certificates, and Background

The main page does not disclose pricing, payment methods, certificates, or accreditation information, nor does it clearly mention institutional backing. In terms of instructor background, the only information is the author’s own note: years of scattered code were organized into a course, and in the third edition AI was used to polish the text, algorithms, and Companion. As a result, its credibility mainly comes from the course structure, open-source repository, and hands-on design rather than formal certification.

Pros, Cons, and Who It’s For

Its strengths are broad and up-to-date coverage plus a high density of practical work, making it suitable for learners who genuinely want to reproduce reinforcement learning algorithms. The downside is that the barrier to entry is not low: the page explicitly recommends a foundation in Python, NumPy, PyTorch, calculus, linear algebra, and probability. It is not very friendly to complete beginners. It is best suited to graduate students, AI engineers, algorithm learners, or those preparing to enter areas such as RLHF, robotics, or world models.

Access from China and Alternatives

The main page does not describe access conditions from mainland China. If the course relies on GitHub, Docker images, or services such as OpenAI, Anthropic, or Gemini, the actual experience may be affected by network conditions. The Companion’s bring-your-own LLM model may also require additional network and payment preparation in China. Comparable alternatives include OpenAI Spinning Up, Hugging Face Deep RL Course, DeepMind/UCL RL courses, and reinforcement learning courses on Coursera/edX.