Matt Strong

Matt Strong

CS PhD at Stanford University

← Paper Digest

Reactive Diffusion Policy: Slow-Fast Visual-Tactile Policy Learning for Contact-Rich Manipulation

Han Xue, Jieji Ren, Wendi Chen, Gu Zhang, Yuan Fang, Guoying Gu, Huazhe Xu, Cewu Lu

Seminal March 2025

The three tasks: peeling a cucumber, wiping a vase and lifting a paper cup with two arms.

TL;DR A slow latent diffusion policy plans action chunks at 1-2 Hz while a fast decoder adjusts them from tactile or force feedback at 24 Hz, trained on demos from an AR teleoperation system that shows touch. On peeling, wiping and bimanual lifting it beats Diffusion Policy by a wide margin (peeling score 0.95 vs 0.44).

What they do

Imitation-learning policies with action chunking, such as Diffusion Policy and ACT, execute each chunk open-loop, so they cannot react to touch mid-chunk. The paper adds two pieces: TactAR, a teleoperation system that shows tactile / force feedback in AR on a Meta Quest 3, and Reactive Diffusion Policy (RDP), a slow-fast visual-tactile policy.

How it works

  • TactAR. The marker displacement field of an optical tactile sensor, or a force sensor’s wrench, is drawn as 3D arrows on the end-effector in AR. The headset costs $499.
  • Fast policy. An asymmetric tokenizer: a 1D-CNN encoder compresses an action chunk into a latent, and a GRU decoder rebuilds the actions from the latent plus the current tactile features (PCA of the marker field, or the 6-D wrench). Only the decoder sees touch, so it carries the reactive corrections. It runs at 24 Hz in under 1 ms.
  • Slow policy. A latent diffusion policy predicts latent action chunks at 1-2 Hz from images, touch and proprioception.
  • Relative end-effector trajectories and latency matching (dropping predicted steps already in the past) both matter for performance.

Results

  • Overall scores on Peeling / Wiping / Bimanual Lifting: Diffusion Policy 0.44 / 0.57 / 0.00, RDP with force input 0.95 / 0.87 / 0.70. Adding tactile embeddings to Diffusion Policy did not help (0.39 / 0.50 / 0.08).
  • When the cucumber is moved after contact, RDP (GelSight) scores 0.80 against 0.15 for Diffusion Policy with tactile embeddings. GelSight and MCTac give similar Peeling scores (0.90 vs 0.88).
  • Ablations: on Peeling, keeping only normal force lowers RDP from 0.95 to 0.48, a symmetric tokenizer to 0.58. On Wiping, shrinking Diffusion Policy’s chunk from 8 to 2 drops its grasp rate from 100% to 20%.
  • TactAR user study (10 users, 200 trials): peel length rises from 0.72 to 0.91 and the share of stable contact force from 0.58 to 0.87.

Limitations

AR feedback is less intuitive than working with one’s own hands. Only two-finger grippers are supported, the fast policy reacts to touch but not to high-frequency vision, and each policy learns a single task.

Videos

Peeling while a person pushes the cucumber down to perturb the contact.
TactAR renders force-sensor, GelSight and MCTac readings as arrows in AR.

Figures

From Figure 1: TactAR, a low-cost teleoperation system that shows tactile / force feedback in AR, and Reactive Diffus...
From Figure 1: TactAR, a low-cost teleoperation system that shows tactile / force feedback in AR, and Reactive Diffusion Policy, which models complex trajectories with a slow policy and closes the loop on high-frequency tactile / force feedback with a fast policy.
From Figure 6: (a) Training: first the fast policy (asymmetric tokenizer), then the slow latent diffusion policy. (b)...
From Figure 6: (a) Training: first the fast policy (asymmetric tokenizer), then the slow latent diffusion policy. (b) Inference: the slow policy predicts a latent action chunk from low-frequency observations; the fast policy decodes it step by step using high-frequency tactile / force input.
From Figure 10: Ablation on Peeling with perturbation after contact. Predicting absolute actions or removing latency ...
From Figure 10: Ablation on Peeling with perturbation after contact. Predicting absolute actions or removing latency matching scores far below full RDP (MCTac).
Abstract

Humans can accomplish complex contact-rich tasks using vision and touch, with highly reactive capabilities such as fast response to external changes and adaptive control of contact forces; however, this remains challenging for robots. Existing visual imitation learning (IL) approaches rely on action chunking to model complex behaviors, which lacks the ability to respond instantly to real-time tactile feedback during the chunk execution. Furthermore, most teleoperation systems struggle to provide fine-grained tactile / force feedback, which limits the range of tasks that can be performed. To address these challenges, we introduce TactAR, a low-cost teleoperation system that provides real-time tactile feedback through Augmented Reality (AR), along with Reactive Diffusion Policy (RDP), a novel slow-fast visual-tactile imitation learning algorithm for learning contact-rich manipulation skills. RDP employs a two-level hierarchy: (1) a slow latent diffusion policy for predicting high-level action chunks in latent space at low frequency, (2) a fast asymmetric tokenizer for closed-loop tactile feedback control at high frequency. This design enables both complex trajectory modeling and quick reactive behavior within a unified framework. Through extensive evaluation across three challenging contact-rich tasks, RDP significantly improves performance compared to state-of-the-art visual IL baselines. Furthermore, experiments show that RDP is applicable across different tactile / force sensors. Code and videos are available on https://reactive-diffusion-policy.github.io.

Summaries are written automatically from each paper's text. Figures and videos come from the paper and its project page; all credit goes to the authors.