Part of a public week-one log of moving from ML/quant work into robotics: what was
actually run, what it cost, and the honest next steps. The scripting was done with an
AI coding agent; every number below was measured on one CPU-only Linux box (Python 3.12,
MuJoCo 4.0.1, PyTorch 2.15 CPU wheels).
Why MuJoCo first, not VLA
The robotics job market talks about VLA and foundation models, but the
interviews keep probing the layer underneath: state estimation, kinematics,
contact, and the discipline of measuring before claiming. MuJoCo is free, runs on
a laptop CPU, and every experiment below is checkable by a reader in under an
hour.
Week one, day by day
- Day 0 — environment. Locked-down Linux, no GPU, PEP 668 system Python.
pip install --target+PYTHONPATH+ CPU torch wheels. Cost: about an hour, most of it reading error messages. The recipe is written down so it costs you ten minutes. - Day 1 — measure before training.
env.step()on HalfCheetah: ~50 microseconds. So the bottleneck is gradient steps, not physics: a SAC update at batch 256 costs about 6 ms. Multiply before you commit. - Day 1 — Pendulum sanity check. The full pipeline (squashed Gaussian, twin critics, learned temperature) reaches roughly -110 to -170 return by 100k steps on CPU (measured with both one and two gradient updates per step). Any RL code you write should pass this before you touch expensive tasks.
- Day 1-2 — HalfCheetah at paper budget. 1M steps, one seed: 109 minutes wall clock on 16 cores. Final evaluation 9,180 (best 9,914) against the SAC paper's ~6,000 reference for the same budget. The evaluation curve, logs and the comparison table are all generated files — the habit that separates an experiment from a screenshot.
- Day 2 — write it up. A reproduction README with "what is reproduced / what is not claimed" is a portfolio artifact that interviewers can verify. This is the piece most portfolios skip.
What a hiring manager can verify
Everything above ships with logs: eval.csv per run, a plot script, and a README
whose tables are generated. The claim is never "I know robotics"; the claim is
"here is a working baseline, here is what it cost, check it yourself".
Next weeks
- Week 2: manipulation — grasping metrics and a pick pipeline in MuJoCo (Menagerie models), CPU-feasible.
- Week 3: a small learning-for-control experiment with a published reference to compare against.
- Week 4: dataset tooling (LeRobot format) — the data-engineering half of the stack.
Everything above is checkable: code, full training logs and the generated
comparison table are MIT-licensed at sac-mujoco-repro. If you'd rather skip
re-running the 1M-step training yourself, the packaged version — logs, learning
curves, checkpoints, README — is on our store for $1. The price is there
to collect first reviews, not to gate the free repo.


