winstonjq/embodied-interview-qa

This repository is a Chinese-language interview question bank for engineering roles in embodied AI, covering humanoid and quadruped robot policies, vision-language-action models, imitation learning, reinforcement learning, world models, and the work to put them on real hardware. The README opens with an English subtitle and describes 2024 to 2026 as a hiring boom for these jobs in China, with questions scattered across multiple Chinese forums. The project collects questions that showed up at least three times across sources into one place. The site is hosted on GitHub Pages and renders as a static HTML page that opens in any phone or desktop browser. Questions default to collapsed, using HTML5 details elements with zero JavaScript, so the reader can think through an answer before clicking to reveal it. Each answer is capped at around 350 characters and is followed by a one-line common mistake note. The README states the project is meant as a quick refresher before an interview, not a long-form tutorial: derivations and code blocks are left out. The bank is split into six volumes. Volume one covers basics like deep learning, intro RL, and robotics with 44 questions. Volume two has 40 questions on RL algorithms including PPO, SAC, TD3, offline RL, and RLHF. Volume three has 58 questions on VLA and imitation learning, covering OpenVLA, RT-2, pi models, and Diffusion Policy. Volume four has 31 questions on world models and sim-to-real, including Dreamer, V-JEPA, domain randomization, and Isaac Lab. Volume five has 39 questions on engineering topics like deploying VLA on Jetson and teleoperation data flywheels. Volume six has 48 questions on legged robot control and teleoperation. The main tables hold 257 questions, with about 16 lower-frequency questions kept as backup at the end of each volume. Each question is tagged L1, L2, or L3 by difficulty. Answers were written by an AI agent and then reviewed by a different AI to lower the chance of single-model errors. The README closes with a long write-up of how the bank was made, calling it a 'vibe coding' project: Claude Code (Opus 4.7 with a 1M context) acted as the controller and dispatched subagents that each completed one volume end to end, while a separate Codex instance acted as the cross-model reviewer through MCP. The project is MIT licensed and contributions of new questions are accepted by issue or pull request, provided each question comes from a public source.