CoRL 2026 · Paper 1930

Robot-DIFT
Correspondence-Sensitive Diffusion Features
for Contact-Rich Robot Manipulation

Yu Deng1,4*Yufeng Jin1,6*Xiaogang Jia2Jiahong Xue1Gerhard Neumann2,3Georgia Chalvatzaki1,4,5

1 TU Darmstadt2 KIT3 FZI4 Hessian.AI5 Robotics Institute Germany6 Honda Research Institute Europe GmbH

* Equal contribution

Paper ↗Code ↗Video ↓Poster ↗OpenReview ↗
TU Darmstadt Karlsruhe Institute of Technology FZI Hessian.AI Robotics Institute Germany Honda Research Institute Europe

Precise control requires visual features that preserve small, action-relevant changes.
Robot-DIFT adapts diffusion correspondence features to robot data and exposes their spatial detail through a multi-scale readout.

2:30 research overview and robot demonstrations · Silent video

Why local correspondence matters

A policy can recognize the correct object yet miss the final alignment. A pin just left or right of a hole requires opposite corrections.

We call the required property correspondence sensitivity: nearby object configurations that require different actions should produce local, discriminative feature changes.

How can diffusion correspondence features become a reusable representation for precise robot control?

Two objects with five annotated surface points and cosine-similarity heatmaps comparing Robot-DIFT, DINOv3 and SigLIP
Local point discrimination. Lower off-diagonal similarity indicates more distinct features at different surface points. This diagnostic complements the downstream manipulation evaluation.

Robot-DIFT overview

01 / Backbone pretraining

DROID pretraining with Manifold Distillation

Imitation learning adapts the Student to robot observations and actions. Annealed feature alignment to a frozen diffusion Teacher helps retain the pretrained spatial features.

Manifold Distillation: clean latent inputs pass through a Student and noisy latent inputs through a frozen Teacher; a projection head aligns their features
Teacher–Student alignment follows CleanDIFT. Robot-DIFT combines this alignment with imitation learning on DROID and gradually reduces the alignment weight during pretraining.

02 / Policy-facing features

S2-FPN combines coarse semantics and fine spatial detail

A Spatial–Semantic Feature Pyramid Network fuses three Student decoder scales before policy pooling. Downstream tasks reuse the frozen Student and train a new readout and policy.

S2-FPN combines coarse-to-fine visual features with language information before the action policy
The diffusion Teacher and alignment heads are removed after pretraining. The Student requires a single forward pass; the action policy generates an action chunk.

Evidence across two manipulation benchmarks

RoboCasa compares visual representations under a shared policy setup. LIBERO-10 evaluates language-conditioned transfer with a frozen Student.

RoboCasa

Average task success (%)
MethodSuccess ↑
CLIP19
SigLIP27
VGGT22
FastVGGT23
DINOv228
DINOv333
SiT29
DIFT38
Robot-DINO40
Robot-DIFT48 ± 1

LIBERO-10

Average task success (%)
MethodSuccess ↑
DP-C53
DP-T58
OpenVLA54
π085
π0-FAST60
UniPi0
UVA90
DIFT-VLA68
Robot-DIFT92 ± 1

Robot-DIFT: three-seed mean ± standard deviation; baselines: one seed. 50 rollouts per task. Baseline sets differ across the two benchmarks. Values follow Tables 1 and 3(a) and the accompanying text in the paper.

Both the backbone and the readout matter. Controlled backbone/readout experiments show that S2-FPN improves policy performance, while diffusion-derived features retain an advantage under the same readout. The full paper includes these controls, correspondence tests and real-robot validation.

Read the complete experiments and protocols →

Abstract

Robot manipulation often fails in the final millimeters: a policy may recognize the right object yet miss the pose offsets, boundaries, or pre-contact alignments needed for action. We argue that such failures arise when semantic invariance suppresses correspondence cues for closed-loop control, or when these cues are not exposed to the policy in a usable form. Modern visual encoders provide strong semantic abstractions, but contact-rich manipulation requires correspondence sensitivity: discriminative feature responses to action-relevant changes in pose, boundary, and contact geometry. Diffusion features provide a strong prior for dense correspondence, but direct use is impractical due to stochasticity, latency, and representation drift. We introduce Robot-DIFT, a deterministic diffusion-derived backbone for real-time control. Through Manifold Distillation, Robot-DIFT converts a noise-conditioned diffusion Teacher into a clean-input, single-pass Student while preserving the teacher’s feature manifold. A Spatial–Semantic Feature Pyramid Network (S2-FPN) fuses coarse-to-fine Student decoder features into visual tokens that expose semantic context and fine contact detail to the policy. Across RoboCasa, LIBERO-10, and real robots, Robot-DIFT outperforms vision–language, self-supervised, geometry-oriented, and diffusion baselines on contact-sensitive tasks. Controlled backbone/readout swaps show that S2-FPN unlocks, rather than replaces, the diffusion correspondence prior.

BibTeX

@inproceedings{deng2026robotdift,
  title = {Robot-DIFT: Correspondence-Sensitive Diffusion Features
           for Contact-Rich Robot Manipulation},
  author = {Deng, Yu and Jin, Yufeng and Jia, Xiaogang and Xue, Jiahong
            and Neumann, Gerhard and Chalvatzaki, Georgia},
  booktitle = {Conference on Robot Learning (CoRL)},
  year = {2026},
  url = {https://arxiv.org/abs/2602.11934}
}