Research map / Image generation / Autoregressive and latent generation
Diffusion-EDFs: Bi-equivariant Denoising Generative Modeling on SE(3) for Visual Robotic Manipulation
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes Diffusion-EDFs, an SE(3) bi-equivariant diffusion generative policy on point clouds that learns visual manipulation from 5-10 demonstrations in under an hour.
Problem
Visual robotic manipulation from stochastic human demonstrations needs data-efficient, generalizable 6-DoF policies; prior diffusion SE(3) methods need many demos and EDFs are slow to train (>10 hours) and have small receptive fields.
Approach
Combines SE(3) bi-equivariant score matching on the Lie group with a Brownian diffusion kernel, a contact-based diffusion origin selection mechanism, and a U-Net-like multiscale Equiformer-based EDF score model, enabling fast end-to-end training without pretraining or object segmentation.
Key results
- Diffusion-EDFs achieves about 0.97 total success on mugs and 0.83 on bottles in the default simulated setting, while R-NDFs and SE(3)-Diffusion Fields get 0.00 without object segmentation.
- Diffusion-EDFs maintains total success around 0.79-0.87 under combined unseen instances, poses, and clutter scenarios.
- Training takes 20-45 minutes on an RTX 3090, claimed ~15x faster than EDFs (over 10 hours).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use Diffusion-EDFs instead of R-NDFs | Simulation pick-and-place · Unseen instances, poses, and clutters, mug | total success rate: 0.53 -> 0.79 | Helped |
| Use Diffusion-EDFs instead of SE(3)-Diffusion Fields | Simulation pick-and-place · Unseen instances, poses, and clutters, mug | total success rate: 0.03 -> 0.79 | Helped |
| Use Diffusion-EDFs instead of R-NDFs | Simulation pick-and-place · Unseen instances, poses, and clutters, bottle | total success rate: 0.72 -> 0.87 | Helped |
| Use Diffusion-EDFs instead of SE(3)-Diffusion Fields | Simulation pick-and-place · Unseen instances, poses, and clutters, bottle | total success rate: 0.00 -> 0.87 | Helped |
| Replace EDF energy-based training with diffusion-based training | Diffusion-EDFs · Training time | training speed: ×15 faster | Helped |
Surprising findings
- R-NDFs and SE(3)-Diffusion Fields achieve zero success on both mug and bottle tasks when object segmentation is removed.
Limitations
- Cannot perform control-level or trajectory-level inference.
- Requires a grasp observation procedure, preventing closed-loop inference.
- Real-world bowls and bottles tasks depend on a separately scanned grasp point cloud.
Experimental setup
Datasets: human demonstrations (10 per task in simulation; 4-10 real)
Benchmarks: simulated mug and bottle pick-and-place under unseen instances, poses, clutters
Metrics: pick-and-place success rate
Baselines: R-NDFs SE(3)-Diffusion Fields
Builds on: Equivariant Descriptor Fields SE(3)-Diffusion Fields Equiformer Annealed Langevin MCMC
Closest papers by meaning
- ET-SEED: EFFICIENT TRAJECTORY-LEVEL SE(3) EQUIVARIANT DIFFUSION POLICY (ICLR 2025 · similarity 0.90)
- Two-Steps Diffusion Policy for Robotic Manipulation via Genetic Denoising (NeurIPS 2025 · similarity 0.90)
- RDT-1B: a Diffusion Foundation Model for Bimanual Manipulation (ICLR 2025 · similarity 0.88)
- Prediction with Action: Visual Policy Learning via Joint Denoising Process (NeurIPS 2024 · similarity 0.88)
- Latent Diffusion Planning for Imitation Learning (ICML 2025 · similarity 0.88)
- VidMan: Exploiting Implicit Dynamics from Video Diffusion Model for Effective Robot Manipulation (NeurIPS 2024 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.