Ambient Diffusion Update

Jan 21, 2026

Adam Wei

Agenda

1. Book keeping items

2. Real-world results

3. Motion planning experiments

4. Conditioning and ambient in the multi-task setting

Book-keeping Items

  • Planned to submit RSS... bailed on deadline
  • Goal is to arxiv by end of Feb, submit to Corl in Apr/May

Distribution Shifts in Robot Data

  • Sim2real gaps
  • Noisy/low-quality teleop
  • Task-level mismatch
  • Changes in low-level controller
  • Embodiment gap
  • Camera models, poses, etc
  • Different environment, objects, etc

robot teleop

simulation

Open-X

Immediate Next Direction

  • Sim2real gaps
  • Noisy/low-quality teleop
  • Task-level mismatch
  • Changes in low-level controller
  • Embodiment gap
  • Camera models, poses, etc
  • Different environment, objects, etc

Open-X

Immediate Next Direction

Open-X

Variant of Ambient Omni

Cool Task!!

Cool Demo!!

Open-X Embodiment

Open-X

Magic Soup++: 27 Datasets

Custom OXE: 48 Datasets

  • 1.4M episodes
  • 55M "datagrams"

Table Cleaning

Ambient policy on OOD objects. (2x speed)

Table Cleaning: Evaluation

Task completion =

   0.1 x [opened drawer]

+ 0.8 x [# obj. cleaned / # obj.]

+ 0.1 x [closed drawer]

Question: How to compute error bars?

\(\sigma=0\)

\(\sigma=1\)

"Clean" Data

Clean Only

Co-train

\(\sigma=0\)

\(\sigma=1\)

"Corrupt" Data

"Clean" Data

... with some mixing ratio \(\alpha\)

Ambient

\(\sigma=0\)

\(\sigma=1\)

\(\sigma_{min}\)

\(\sigma > \sigma_{min}\)

"Clean" Data

"Ambient"

Locality

\(\sigma=0\)

\(\sigma=1\)

\(\sigma_{max}\)

\(\sigma \leq \sigma_{max}\)

"Clean" Data

"Locality"

Ambient + Locality

\(\sigma=0\)

\(\sigma=1\)

\(\sigma_{max}\)

\(\sigma \leq \sigma_{max}\)

"Clean" Data

"Locality"

\(\sigma > \sigma_{min}\)

"Ambient"

\(\sigma_{min}\)

Ambient [OLD SLIDE]

Even the best policy right now is mediocre...

Results: 50 Clean Demos

Results: 150 Clean Demos

Results: 150 Clean Demos

Best 50 demo policy: 83.0%

150 clean demos: 80.8%

Ex. Reweights vs Unweighted

Reweighted

(sample clean 50%)

Unweighted

(sample clean 0.06%)

Ablations: Unweighted vs Weighted

Ablations: Ambient Loss vs Denoising Loss

Ablations: Noisy Observations

\(\sigma=0\)

\(\sigma=1\)

\(\sigma_{min}\)

\(\sigma > \sigma_{min}\)

"Clean" Data

"Ambient"

Noise observation w.p. \(p\)

Produce model with noise level of observation

Ablations: Noisy Observations

\(\sigma=0\)

\(\sigma=1\)

\(\sigma_{min}\)

\(\sigma > \sigma_{min}\)

"Clean" Data

"Ambient"

Noise observation w.p. \(p\)

Produce model with noise level of observation

Inference time: set noise level of observations to 0

Ablations: Noisy Observations

Motion Planning

Distribution shift: Low-quality, noisy trajectories

High Quality: 

100 GCS trajectories

Low Quality: 

5000 RRT trajectories

Loss Function

Loss Function (for \(x_0\sim q_0\))

Denoising Loss vs Ambient Loss

Choosing \(\sigma_{min}\)

Motion Planning Results

GCS

Success Rate

(Task-level)

Avg. Jerk^2

(Motion-level)

RRT

GCS+RRT

(Co-train)

GCS+RRT

(Ambient)

50%

Policies evaluated over 100 trials each

100%

7.5k

17k

91%

14.5k

98%

5.5k

Loss Function

Loss Function (for \(x_0\sim q_0\))

Denoising Loss vs Ambient Loss

Choosing \(\sigma_{min}\)

Motion Planning: 7-DoF Robot Arms

Generate good (expensive) and bad (cheap) motion planning data in 20,000 environments

Goal: generate high-quality, collision-free trajectories in new environments

Loss Function

Loss Function (for \(x_0\sim q_0\))

Denoising Loss vs Ambient Loss

Choosing \(\sigma_{min}\)

Too hard :(

Set up

  • 10k trajopt trajectories
  • 100k shortcut + RRT trajectories
  • 6 RGB views

 

Results

  • Doesn't work, tons of collisions

Loss Function

Loss Function (for \(x_0\sim q_0\))

Denoising Loss vs Ambient Loss

Choosing \(\sigma_{min}\)

Experiment Goal

Primary Goal: To provide intuition for ambient diffusion in a simple environment

  • Effect of  \(\sigma_{min}\)
  • Ambient loss
  • Learning 
  • Generating good trajectories from bad data: i.e. learn only the task structure from bad data

Secondary Goal: General neural motion planner

  • There are entire papers on this...

Loss Function

Loss Function (for \(x_0\sim q_0\))

Denoising Loss vs Ambient Loss

Choosing \(\sigma_{min}\)

Memorize Single Environment

Clean data:

  • 100k trajopt trajectories

Corrupt data

  • 500k RRT trajectories
  • 500k RRT + shortcut trajectories

Observation space: proprioception, target joint position

Action space: joint position commands

Loss Function

Loss Function (for \(x_0\sim q_0\))

Denoising Loss vs Ambient Loss

Choosing \(\sigma_{min}\)

[Preliminary] Results

Trajopt

Success Rate

(Task-level)

Avg. Length

(Motion-level)

RRT

Trajopt + RRT

(Co-train)

Trajopt+RRT(Ambient)

30.9%

Policies evaluated over 1000 trials each

?

5.97

?

24.1%

Avg. Acc^2

(Motion-level)

2.75

?

11.24

7.32

38.5%

6.72

3.57

Loss Function

Loss Function (for \(x_0\sim q_0\))

Denoising Loss vs Ambient Loss

Choosing \(\sigma_{min}\)

Next Steps

  1. Answer to Russ' multi-task question
  2. Remaining experiments
    1. Finish neural motion planning
    2. LBM Eval
    3. Lu's cross embodiment data
    4. Tower-building task (real-world)
  3. Write paper

Loss Function

Loss Function (for \(x_0\sim q_0\))

Denoising Loss vs Ambient Loss

Choosing \(\sigma_{min}\)

Cross Embodied Data

Loss Function

Loss Function (for \(x_0\sim q_0\))

Denoising Loss vs Ambient Loss

Choosing \(\sigma_{min}\)

Cross Embodied Data

Amazon CoRo 01/21/26

By weiadam

Amazon CoRo 01/21/26

  • 7

More from weiadam

  • RLG Fall 2025 Long Talk: Learning From Low-Quality Data
    weiadam
    174
  • Unifying Score-Based and Denoiser-Based Generative Modeling
    weiadam
    357