FixTheRobot Explore Robotics / Robot Learning / M11 / M11-L2

This lab works on phones, but writing code is much easier on a laptop.

Robot Learning · Research Practice and Capstone · M11-L2 · Advanced

The Learning Capstone

Experiment log · Scout at Meridian General Hospital

Experiment log, Meridian General Hospital, facilities office, Friday. The night-shift lead has asked one question before Scout is allowed to deliver on its own: does the learned policy reach the goal reliably, from every part of the hall, or only on average? Nobody wants a slide with one big number. They want to know where it fails, how it fails, and how sure you are.

You will run the whole project on Scout's reach task. First train a policy from the expert's demonstrations and make it work from every start region. Then evaluate a finished experiment by region with honest intervals, diagnose six failed runs from their metrics by a written rule, and finally assemble the report from three seeds. The page trains and deploys the policies, measures everything on hidden starts, and checks your report against a rubric.

Your task

    Simulation
    solution.pyLoading Python…
    Ctrl+Enter runs
    What Python can I use?

    Plain Python with math and heapq. Your code runs in a background worker: an endless loop is stopped after a few seconds and the page stays usable. It runs in your browser and is never sent anywhere.

    Hints

    What this lab teaches

    Stage 1 is behaviour cloning end to end: choose features, fit them by ridge regression, and deploy your weights on hidden starts. It is graded per start region (easy, mid, near), because a policy that fails near the pillar can still have a good average. Stage 2 turns episode records into per-region success rates with 95 percent Wilson intervals and names the weakest region by rate; dividing a region by the total number of episodes is the classic slip. Stage 3 applies a short diagnostic rule to the evidence (training versus evaluation success, how runs fail, which regions the demonstrations cover) and separates underfitting, a missing feature, a distribution shift and too few demonstrations, including the run that has nothing wrong. Stage 4 assembles the experiment report from three seeds.

    Grading is result-based on hidden policies, demonstrations and starts. The report is checked by a rubric a program can apply: the numbers match the measured values within 0.005, the weakest region and how it fails are named, the number of seeds and episodes is stated, at least one limitation is given, and the claim is no stronger than the lower end of the interval it rests on. Pointing out what a study cannot show is part of the result (lessons M11-T3 to M11-T5).

    Theory for this lab

    Research Practice and Capstone labs

    1. M11-L1: Compare Two Policies Statistically
    2. M11-L2: The Learning Capstone: this lab.

    Educational simplification: one simulated task and a noise-free expert, linear policies deployed on three known feature maps, three start regions, a few dozen episodes per region, thresholds (0.9, 0.8, 0.1) that are this lab's own acceptance limits, and pooled episodes across three seeds treated as one sample. All data are simulated; nothing here is evidence about a physical robot.

    Get the next lab

    Get notified when new robotics labs are released. One email per new lab, nothing else.