Lesson 19 of 21 in this unit
Vary the world during training.
Compare randomized collection with separately changed evaluation conditions.
By the end: Compare randomized collection with separately changed evaluation conditions.
Prerequisites and path options
Why learn this? Find which simulated variations help a policy and which deployment changes remain outside its training evidence.
Start with something concrete.
A robot taught on one smooth table may move too little on a rough surface. Showing several tables can help, but only for effects the simulator actually represents.
Try the interactive modelCan randomizing drag teach an effect absent from every training model?
Change one thing. Watch what follows.
Loading this browser experiment…
Understand what changed.
A shorter explanation, with the original worked examples available below. Both explain the same experiment and its limits.
The question
A robot taught on one smooth table may move too little on a rough surface. Showing several tables can help, but only for effects the simulator actually represents.
Try it and inspect what changed
Reset and train fixed versus randomized collection with equal updates; compare delay and noise.
Compare fixed and randomized checkpoints at a matched update budget. Keep failure cases as well as successes. In the project freeze the candidate before opening the separate scenario combinations. Reusing those combinations later is development practice, and known exposure stays recorded.
The calculation in words
Randomized collection varies drag over 0, 0.2, 0.5 and 0.7; noise changes the delivered coordinates and delay selects earlier observations. The teacher action still uses its contemporaneous state. That mismatch makes delayed/noisy imitation a genuine harder mapping. Object radius is recorded too, but these demonstration paths do not touch walls.
Keep the conclusion within the evidence
Changing a parameter name is not sufficient intervention. Since radius does not affect teacher actions away from walls, its randomization adds no collision-avoidance examples here. State this limitation in the data card. A stronger study would collect meaningful boundary interactions under a contact-capable simulator.
Drag is an authored motion attenuation coefficient, not a Coulomb friction law. Neither distribution covers sticky contact, rotation or actuator lag. A broader training range can also worsen in-range precision; measure the tradeoff instead of assuming robustness.
You can explore every lesson directly. Completing this page, submitting a project report and demonstrating understanding on a separate case are different records. Optional code is available in the robot-learning studio; it does not increase guided lesson completion requirements.
Open the full worked examples, calculations and limitations
This is the same reference used in Technical mode. Expand it whenever you need the exact numbers, assumptions or further explanation.
Start with the task
A robot taught on one smooth table may move too little on a rough surface. Showing several tables can help, but only for effects the simulator actually represents.
Mechanism and units
Randomized collection varies drag over 0, 0.2, 0.5 and 0.7; noise changes the delivered coordinates and delay selects earlier observations. The teacher action still uses its contemporaneous state. That mismatch makes delayed/noisy imitation a genuine harder mapping. Object radius is recorded too, but these demonstration paths do not touch walls.
Changing a parameter name is not sufficient intervention. Since radius does not affect teacher actions away from walls, its randomization adds no collision-avoidance examples here. State this limitation in the data card. A stronger study would collect meaningful boundary interactions under a contact-capable simulator.
Work through the evidence
Compare fixed and randomized checkpoints at a matched update budget. Keep failure cases as well as successes. In the project freeze the candidate before opening the separate scenario combinations. Reusing those combinations later is development practice, and known exposure stays recorded.
Reset and train fixed versus randomized collection with equal updates; compare delay and noise.
What this experiment can establish
Drag is an authored motion attenuation coefficient, not a Coulomb friction law. Neither distribution covers sticky contact, rotation or actuator lag. A broader training range can also worsen in-range precision; measure the tradeoff instead of assuming robustness.
Keep collection, fitting, frozen execution and interpretation as separate records. Save the configuration and complete result, including failures, in the robot-learning project. Report submission checks required evidence and text presence; reasoning and implementation quality need independent review.
Radius is varied, but all teacher demonstrations stay far from walls. What collision skill does that establish?
Take this idea with you.
Drag is an authored motion attenuation coefficient, not a Coulomb friction law. Neither distribution covers sticky contact, rotation or actuator lag. A broader training range can also worsen in-range precision; measure the tradeoff instead of assuming robustness.
Explore related topics and the topic map
How this topic connects
Explore this topic on the interactive mapUse these connections to review a foundation or choose where to go next. You can open any topic without completing the others.
What it builds on
- Geometry and control
Prevent coordinate-frame mistakes when combining robot positions, sensor readings and targets.
Compare randomized collection with separately changed evaluation conditions.
What it enables
- Robot-learning project Apply in a project
Combine demonstration provenance, policy comparisons and complete-task evidence in the robot-learning project; keep embodiment limits and review status explicit.