Lesson 16 of 21 in this unit
Generate more than one valid action plan.
Train conditional action denoising and independently verify sampled routes.
By the end: Train conditional action denoising and independently verify sampled routes.
Prerequisites and path options
Why learn this? Inspect distinct valid action modes and the constraints a generative action model does not enforce.
Start with something concrete.
Two demonstrations go around opposite sides of a post. Their average can go straight through it, even though both examples were successful.
Try the interactive modelWhy can the mean of two valid routes fail?
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
Two demonstrations go around opposite sides of a post. Their average can go straight through it, even though both examples were successful.
Try it and inspect what changed
Train the denoiser, compare noise seeds −1 and +1 and inspect clearance.
At each reverse step compare the noisy input, predicted clean value and next value. A later sample is not necessarily valid. The independent segment-to-circle check includes segment interiors and accepts only strict clearance from a 0.25 m obstacle. Compare the zero-coefficient mean route.
The calculation in words
Each route has thirteen waypoints: x runs from −1 to +1 m and y is a learned coefficient times a sine arch. The clean coefficient has two modes, −0.65 and +0.65 m, shifted by the condition. Forward noise uses sqrt(alpha) times the clean value plus sqrt(1-alpha) times a Gaussian draw. Four alpha values are 0.95, 0.7, 0.4 and 0.15.
Keep the conclusion within the evidence
Six fitted polynomial coefficients predict the clean route coordinate from noisy input, noise level and condition. The deterministic reverse update reuses the inferred noise while moving to the next alpha. This is actual denoising training in a restricted basis. It does not learn arbitrary action sequences, contact dynamics or visual representations.
The Diffusion Policy primary project (RSS 2023 / IJRR 2024, reviewed 23 October 2026) studies conditional action diffusion and receding-horizon control. Our scalar teaching model is not a reproduction of its architecture, benchmarks or hardware results.
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
Two demonstrations go around opposite sides of a post. Their average can go straight through it, even though both examples were successful.
Mechanism and units
Each route has thirteen waypoints: x runs from −1 to +1 m and y is a learned coefficient times a sine arch. The clean coefficient has two modes, −0.65 and +0.65 m, shifted by the condition. Forward noise uses sqrt(alpha) times the clean value plus sqrt(1-alpha) times a Gaussian draw. Four alpha values are 0.95, 0.7, 0.4 and 0.15.
Six fitted polynomial coefficients predict the clean route coordinate from noisy input, noise level and condition. The deterministic reverse update reuses the inferred noise while moving to the next alpha. This is actual denoising training in a restricted basis. It does not learn arbitrary action sequences, contact dynamics or visual representations.
Work through the evidence
At each reverse step compare the noisy input, predicted clean value and next value. A later sample is not necessarily valid. The independent segment-to-circle check includes segment interiors and accepts only strict clearance from a 0.25 m obstacle. Compare the zero-coefficient mean route.
Train the denoiser, compare noise seeds −1 and +1 and inspect clearance.
What this experiment can establish
The Diffusion Policy primary project (RSS 2023 / IJRR 2024, reviewed 23 October 2026) studies conditional action diffusion and receding-horizon control. Our scalar teaching model is not a reproduction of its architecture, benchmarks or hardware results.
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.
Two clear route midpoints are −0.65 m and +0.65 m; the obstacle covers ±0.25 m. Is their mean clear?
Take this idea with you.
The Diffusion Policy primary project (RSS 2023 / IJRR 2024, reviewed 23 October 2026) studies conditional action diffusion and receding-horizon control. Our scalar teaching model is not a reproduction of its architecture, benchmarks or hardware results.
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.
Train conditional action denoising and independently verify sampled routes.
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.