楼主目前计划转行，本身优势是有robotic的本硕，但毕业后种种原因并没有从事相关工作，今年看到领域火热（钱多 
）所以开始自己转行，最终想做人形机器人相关的RL方面工作，目标是年底面试能够进入panel，panel能过不过就不好说了。
本职工作可以说不软件也不硬件，manufacturing 相关
这是Claude给的track，有没有坛友是行业相关帮忙参考一下。 
目前已经复习完基础ML，写完了transformer（上学学过但都忘干净了
以下用的相关job post 生成的，
**M1. PPO from scratch
Single file, ~400 lines, no RL library. Continuous action space, Gaussian policy, GAE, clipped objective, value loss, entropy bonus.
Done when: it solves Pendulum-v1 and reaches ~2500+ on HalfCheetah, and you can whiteboard the policy-gradient derivation and explain what the clip does without notes.
**M2. Reproduce a locomotion baseline in Isaac Lab
Unitree Go2 or ANYmal-C, flat terrain, velocity-command tracking, using rsl_rl’s PPO (not yours — the point here is the env stack, not the algorithm).
Done when: trained policy tracks commanded velocities, converges in <20 min on your 5080, and you have video.
**M3. Rough terrain + curriculum
Terrain curriculum (slopes → stairs → discrete obstacles), domain randomization over friction/mass/latency, push perturbations.
Done when: the policy survives terrain it never saw during early training, and you can show the curriculum progression curve.
M4. Ablation study + writeup (target: late October) ← this is the one that matters
Systematically ablate reward terms and DR ranges. What happens without the action-rate penalty? Without feet-air-time? With friction randomization off? Plot it.
Done when: you have a technical blog post with numbers. This is the single highest-leverage artifact on the list. Anyone can run a tutorial and post a video of a walking robot. Almost nobody can tell you why their reward function is shaped that way and show the evidence. That’s the difference between “followed instructions” and “engineer.”
M5. Humanoid + push recovery (target: December) ← your differentiator
Isaac Lab H1 or G1. Get it walking, then do push recovery — and explicitly compare the learned policy against the classical capture-point / LQR approach
Done when: you can show both, quantify where each wins, and discuss the hybrid. It is exactly the conversation a humanoid locomotion team wants to have, and it makes your “old” background an asset rather than a detour.
