Goal for PhD

• Problem solving
  • Quickly grasp a new problem domain
  • Systematically analyze problems
  • Persistently make progress
• Verbal and writing communication skills
• Knowledge in a specific domain
• Mathematical foundations
• Implementation skills
• Insight and vision

Goal for undergraduates

• Knowledge and skills
• Research experience
• Experience (semi-)PhD life
• Find research taste
• Quickly grasp a new problem domain

Basic skill set

• Go over The Missing Semester of Your CS Education: command-line interface, shell script, zsh, vim, git, ssh, tmux
• Python
  • Use a virtual environment (uv or conda) for each project
  • Use a code formatting tool for easy collaboration: Black is simple yet effective (ruff looks good too)
  • Wrap your code as a package for reusability: How to manage a project code
• English
  • You don’t have to be fluent (I’m not fluent either). But, you should be able to communicate in English, especially speaking.
• Some useful materials: Machine learning, probabilistic graphical model, deep learning or deep learning for computer vision, deep reinforcement learning, reinforcement learning theory, RL implementation, deep unsupervised learning, transformer, deep multi-task and meta learning, artificial intelligence explained by Korean neuroscientist, language modeling from scratch, flow matching and diffusion models
  • Try assignments in JAX (recommended) or PyTorch
• AI tools
  • Use Codex or Claude Code (with Pro or Max subscription)
  • Learn basic prompting/harnessing skills
• Learn LaTeX, a programming language for paper writing, and make your academic CV
• Learn about basic web development (html, css, javascript) and build your academic website

Get ready for research

• RL Implementation

  • Goal: Implement several state-of-the-art methods, do some experiments, and visualize results
  • Try the packages frequently used in our lab: wandb (logging), tqdm (progress bar), matplotlib, seaborn (visualization), hydra-core (configuration), einops (tensor manipulation)
  • Log all experiments and analysis using “reports” in wandb
  • Always run experiments with multiple seeds (at least 3) — One good or bad result from a single seed may mean nothing

• Simulated environments

  • Goal: Understand the code structure of environments and try some of them
  • Gymnasium (previously, OpenAI gym) is the most commonly used interface for RL
  • Gymnasium, mujoco, dm_control, Isaac Lab, IKEA furniture assembly environment, FurnitureSim, and more

• VLA implementation

  • Goal: Understand the code structure of VLA methods and datasets
  • openpi (JAX) and LeRobot (PyTorch) are the most widely used VLA backbone and dataset infra

• 👍 Read papers from our lab

• Read relevant papers

  Robot learning reading list

• Think about problems you want to solve and make an idea slide deck

  Thinking about a research problem

Start your lab life

• Be proactive, participate in meetings, ask questions, think about problems even if not explicitly asked for
• Actively work to get hands-on experience in all parts of the research pipeline: build slide decks to outline ideas, present in group meetings, write paper idea abstracts, help in implementing prototypes and running experiments
• Ask questions sooner rather than later when you get stuck