The situation
The startup (around 80 staff, customers across Japanese SMB electronics manufacturing) charges per integration; cutting the programming hours per new task from 8 to 3 would unlock a long-tail customer segment.
Your task
Train a behavior cloning policy for pick-and-place that achieves competitive success on held-out scenarios versus a hand-scripted baseline.
The brief
You start with
You receive 200 human teleoperated demonstrations (state + action trajectories) of picking 8 small electronic components from a tray and placing them at marked locations in a robosuite simulator.
What you do
- Train a behavior cloning (BC) policy with a small MLP on (state, action) pairs.
- Evaluate on 50 held-out pick scenarios with varied component positions.
- Score on (a) task success rate per component type, (b) average completion time vs. the scripted baseline, (c) graceful failure rate (drop without crash) on out-of-distribution positions.
Success looks like
Overall success above 70 percent and time-to-completion within 1.3x of the scripted baseline.
What you will practice
- Implement behavior cloning from human demonstrations
- Evaluate imitation policies on held-out manipulation scenarios
- Detect and report out-of-distribution behavior
- Communicate imitation-learning trade-offs to a non-ML applications team
A suggested way in
- Normalize state and action vectors before training to stabilize learning
- Train with multiple seeds; the variance is the headline result, not the best run
- Always report per-component success — averaged success hides task-specific failures
Skills you will practice
Behavior cloning, Imitation learning, Manipulation, Policy evaluation, Pytorch, Simulation
Where it leads
Machine Learning Engineer, Applied AI Scientist, ML Researcher
How it is assessed
- Task success35%
Overall success rate above 70 percent on the 50-scenario holdout.
- Evaluation rigor25%
Multiple seeds, per-component breakdown, OOD positions all reported.
- Failure-mode honesty20%
Graceful failure rate is measured and named, not hidden.
- Feasibility memo20%
Memo states honestly whether BC saves programming time at the supplier's quality bar.
Resources
- PyTorchTool
Train the BC policy and run inference.
- robosuiteTool
Manipulation simulator with pre-built pick-and-place environments.
- PythonTool
Glue for training, evaluation, and reporting.
- Weights & BiasesTool
Track training curves across seeds.
- pandasTool
Per-component success-rate bookkeeping.
- NotionTool
Author the feasibility memo.
Finish it. Earn this.
A verified certificate with your name. Anyone can check it.
Time to practice.
Solve this challengeMore like this.
- Software EngineerBuild a Multi-Level Feedback Queue Scheduler in xv6
- Software EngineerReverse-Engineer a Stripped Binary to Restore Linking Symbols
- Software EngineerImplement a Distance-Vector Routing Algorithm
- AI EngineerEnd-to-End Lane Following on a Donkeycar Platform
-
Build a Multi-Level Feedback Queue Scheduler in xv6 Software Engineer Fork the xv6-riscv repository and implement MLFQ in proc.c with 3 priority queues, per-queue time slices (10ms / 20ms / 40ms), priority demotion on sl operating systemsschedulingc programming Open challenge → -
Reverse-Engineer a Stripped Binary to Restore Linking Symbols Software Engineer Receive the stripped shared object (Linux ELF, x86-64, ~280KB), a crash dump with 12 frames of indecipherable addresses, and 4 years of partial git hi reverse engineeringlinking and loadingmachine code Open challenge → -
Implement a Distance-Vector Routing Algorithm Software Engineer Implement distance-vector routing in Python: each simulated router maintains a distance vector, exchanges vectors with neighbors every 1 second of sim routingdistance vectornetwork simulation Open challenge → -
End-to-End Lane Following on a Donkeycar Platform AI Engineer Use the public Donkeycar Tub dataset (or collect about 30 minutes of driving on the simulator). Train a CNN-policy baseline (the Donkeycar default arc end to end learningimitation learningpytorch Open challenge →