Module Check: Robots That Learn
You have covered the full arc of learning and autonomy in robotics: why hand-coding fails, how supervised learning builds perception, how reinforcement learning discovers control policies, how imitation learning extracts skill from demonstration, how simulation bridges to reality, how the autonomy stack integrates all components, how autonomy levels calibrate human-robot authority, and how safety and verification make learned systems trustworthy. This module check asks you to reason with these ideas across all nine lessons — not merely recall them. The flashcards review core vocabulary; the quizzes demand integrated analysis; the synthesis activity asks you to connect the entire arc.
Flashcards — click each card to reveal the answer
Module Quizzes
A robot is trained using behavioral cloning on 1,000 expert demonstrations of navigating a factory floor. During deployment, it performs well for the first 20 steps but fails increasingly after that. When DAgger is applied for three iterations, performance improves dramatically beyond step 20. What precisely explains this improvement?
An RL-trained robotic arm achieves 95% task success in simulation. After transfer to the real robot, success drops to 40%. A domain-randomized version achieves only 85% in simulation but 83% on the real robot. What principle does this comparison demonstrate?
A Level 3 supervised autonomy robot (fully autonomous, human monitoring) operating in a hospital corridor detects a patient who has fallen. The robot's policy has never seen a fallen person during training. What is the correct behavior under the principles discussed in this module?
A team uses inverse reinforcement learning (IRL) instead of behavioral cloning to train a robot chef from expert cooking demonstrations. What specific advantage does IRL offer for generalization to new recipes not in the demonstrations?
A safety engineer proposes that a medical robot's entire policy be replaced with a formally verified classical controller for deployment in a certified environment. A machine learning engineer argues the learned policy significantly outperforms the classical controller on the target task. How should the team resolve this tension?
The autonomy stack's timescale separation assigns perception to ~30 Hz, planning to ~5 Hz, and control to ~1000 Hz. A student proposes running the entire stack at 1000 Hz to maximize responsiveness. What is wrong with this proposal?
Every lesson in this module is a facet of one argument: giving robots useful autonomy in the real world requires making them learn from data, verifying what they have learned, matching their autonomy to their demonstrated competence, and building safety into the architecture rather than assuming it. The best robot engineers treat capability and safety as inseparable — not competing priorities, but co-requisites for any system worth deploying.
Synthesis: Design Review Board
- This capstone activity asks you to integrate learning, autonomy, safety, and sim-to-real reasoning into a structured design review — the same format used in real robotics companies before a system enters field trials.
- System description: An autonomous robot is proposed for deployment in a public elementary school. It is designed to assist teachers by autonomously distributing materials (worksheets, art supplies) to student desks, collecting completed work, and patrolling corridors during recess to report safety hazards to staff. The robot uses a learned end-to-end policy trained on 10,000 teleoperated demonstration trajectories collected in an adult research lab.
- You are on the design review board. Address each of the following:
- Section 1 — Learning Adequacy Review
- Identify three specific ways the training data (adult lab demonstrations) may fail to represent the actual deployment environment (elementary school with children). For each, describe the specific learned behavior failure it could cause and propose a data collection fix.
- Section 2 — Autonomy Level Review
- The proposed system operates at Level 4 (full autonomy within school hours). For each of the robot's three functions (material distribution, work collection, hazard reporting), assign your recommended autonomy level at launch and justify the difference if any.
- Section 3 — Safety Architecture Review
- The robot will operate around children aged 5-11. Specify four hard safety constraints it must never violate. For each, describe the safety enforcement mechanism: CBF, hardware limit, sensor-based interlock, or organizational procedure.
- Section 4 — Verification Review
- Design a pre-deployment verification protocol. What test scenarios are mandatory? What is the minimum number of trials per scenario? What pass/fail criteria do you require? The school district wants deployment by September. You are reviewing in June. Is this timeline feasible?
- Section 5 — Board Decision
- Based on your review, state your recommendation: approve for deployment, approve with conditions (list them), or reject with required changes. Write a one-paragraph justification that any parent of a student in the school could understand.