PYTHON · NUMPY

What the evidence costs

A paper prints a compute budget as 10 x 8 x 3 x 5 = 1,350 GPU-hours. What should you do before putting 1,350 in your own plan?

Rebuild a published reproduction budget from its own parts, from milliseconds per training step up to GPU-hours, check each line against its printed total, and put an energy ceiling on it.

01Challenge

A NeurIPS 2025 paper on reinforcement learning with action chunking tells you what it would cost to reproduce its main results: 10,350 GPU-hours, built from hours per run times methods times tasks times seeds, and it publishes the milliseconds each training step takes. Rebuild the budget from those parts before you quote it, and say how much energy it can hold at most.

Predict firstA paper prints a compute budget as 10 x 8 x 3 x 5 = 1,350 GPU-hours. What should you do before putting 1,350 in your own plan?

02Model

The last two lessons priced the bench: how much of a training hour the robot was moving, and what a policy's thinking time does to a robot that keeps moving. This one prices the evidence. A robot-learning result is a table of success rates, and every cell of it is several seeds of a full training run. This paper, reinforcement learning with action chunking, trains simulated robot arms on 25 OGBench tasks and 3 robomimic tasks, each run 1M offline steps and then 1M online steps, and it does something most papers do not: it tells you what reproducing its main results would cost. Hours per run times methods times tasks times seeds: 6 x 15 x 25 x 4 = 9,000 GPU-hours on OGBench, 10 x 8 x 3 x 5 on robomimic, printed as 1,350, and 10,350 in all. It also publishes the milliseconds each training step takes on its RTX A5000. That gives you two independent ways into the same budget. From the milliseconds you rebuild one run, 3.62 to 7.53 hours across four methods, which brackets the 6 hours the budget uses. From the factors you rebuild each line, and one of them comes to 1,200 where 1,350 is printed. Neither check needs anything but the paper. Then NVIDIA's own 230 W ceiling for the card turns 10,200 GPU-hours into at most 2,346 kWh on the cards alone.

03Guided practice

1. One run, from the milliseconds. A run is an offline half and an online half, each 1M steps. Multiply each half's milliseconds per step by its steps, add the halves, and convert milliseconds to hours. Check yourself: 1,800 one-second steps in each half is one hour. 2. Compare with what the paper says. It says a run takes around 4 to 7 hours and budgets 6. Your rebuilt runs are timed on one task, so read them as a bracket, not a replacement. 3. Rebuild each budget line. Multiply the four factors printed under each line and set the product beside the printed total. The given cell prints both. 4. Put a ceiling on the energy. The card's maximum power is NVIDIA's figure, not the paper's, and a ceiling, not a measurement. Multiply it by the GPU-hours you rebuilt, and say which total you used.

04Feedback

PASS WHEN You rebuilt a published compute budget from its own parts, checked every line against its printed total, and put a labelled ceiling on its energy.

If your run fails, check:

An answer 3,600 times too large stopped at seconds. Milliseconds to hours is a division by 1000 and then by 3600.
An answer near half the right one counted only one half of the run. Add the offline and online halves.
Define run_hours(ms_per_step, steps) returning the hours one whole run takes.

05Retrieve & space

The rebuilt runs span 3.62 to 7.53 hours and the budget uses 6. Why is it reasonable for a budget to use one number here, and what would you want before trusting it for a different task?
One budget line rebuilds exactly and one does not. What are the two readings of the one that does not, and why does your plan not need to decide between them to stay honest?
The 2,346 kWh is a ceiling on the cards alone. Name two things that would make the true energy of reproducing these results lower, and two that would make it higher.

06Mastery & project

You can take a published compute estimate, rebuild it from its own factors and timings, report where it agrees and where it does not, and convert it into a labelled energy bound.

Find a robot-learning paper you would like to reproduce. Write its budget as hours per run times runs, from whatever it publishes: timings, step counts, seeds, tasks, methods. If it prints a total, rebuild it and say whether it agrees. If it prints none, your rebuilt budget is the first one, and the gap you had to fill with assumptions is worth writing down too.

GO FURTHER · THE LIVE PLAYGROUND

Reinforcement Learning with Action Chunking ↗ · Li, Zhou and Levine, NeurIPS 2025, version 4: the budget in Appendix D.1, the per-step timings in Figure 17, and the evaluation protocol in Appendix D.2.
NVIDIA RTX A5000 ↗ · The card's specification, where the 230 W maximum power comes from.
What a bench hour buys ↗ · The same question asked of a real robot's training hours instead of a simulator's.

← Learning while the policy thinks Back to course →

PYTHON · NUMPY

What the evidence costs

One paper's own budget, rebuilt two ways from its own parts, and a ceiling on its energy.

You read

the arrays and values already in scope

You change

the code you write in each cell

Fixed

the dataset and the checks that grade you