Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Adaptive Compliance Policy: Learning Approximate Compliance for Diffusion Guided Control

Yifan Hou, Zeyi Liu, Cheng Chi, Eric Cousineau, Naveen Kuppuswamy, Siyuan Feng, Benjamin Burchfiel, Shuran Song

TL;DR

The paper tackles the gap in visuomotor policies by enabling robots to learn adaptive, task dependent compliance. It introduces Adaptive Compliance Policy, a diffusion-based framework that predicts both a reference pose and a spatially varying stiffness profile, guided by demonstrations and force/visual data. By modeling a simple yet effective stiffness rule and employing kinesthetic demonstrations, ACP achieves substantial improvements on two contact rich tasks, outperforming fixed stiffness and standard compliant policies. The approach demonstrates practical, scalable learning of compliance for robust manipulation in uncertain, real world scenarios.

Abstract

Compliance plays a crucial role in manipulation, as it balances between the concurrent control of position and force under uncertainties. Yet compliance is often overlooked by today's visuomotor policies that solely focus on position control. This paper introduces Adaptive Compliance Policy (ACP), a novel framework that learns to dynamically adjust system compliance both spatially and temporally for given manipulation tasks from human demonstrations, improving upon previous approaches that rely on pre-selected compliance parameters or assume uniform constant stiffness. However, computing full compliance parameters from human demonstrations is an ill-defined problem. Instead, we estimate an approximate compliance profile with two useful properties: avoiding large contact forces and encouraging accurate tracking. Our approach enables robots to handle complex contact-rich manipulation tasks and achieves over 50\% performance improvement compared to state-of-the-art visuomotor policy methods. For result videos, see https://adaptive-compliance.github.io/

Adaptive Compliance Policy: Learning Approximate Compliance for Diffusion Guided Control

TL;DR

The paper tackles the gap in visuomotor policies by enabling robots to learn adaptive, task dependent compliance. It introduces Adaptive Compliance Policy, a diffusion-based framework that predicts both a reference pose and a spatially varying stiffness profile, guided by demonstrations and force/visual data. By modeling a simple yet effective stiffness rule and employing kinesthetic demonstrations, ACP achieves substantial improvements on two contact rich tasks, outperforming fixed stiffness and standard compliant policies. The approach demonstrates practical, scalable learning of compliance for robust manipulation in uncertain, real world scenarios.

Abstract

Compliance plays a crucial role in manipulation, as it balances between the concurrent control of position and force under uncertainties. Yet compliance is often overlooked by today's visuomotor policies that solely focus on position control. This paper introduces Adaptive Compliance Policy (ACP), a novel framework that learns to dynamically adjust system compliance both spatially and temporally for given manipulation tasks from human demonstrations, improving upon previous approaches that rely on pre-selected compliance parameters or assume uniform constant stiffness. However, computing full compliance parameters from human demonstrations is an ill-defined problem. Instead, we estimate an approximate compliance profile with two useful properties: avoiding large contact forces and encouraging accurate tracking. Our approach enables robots to handle complex contact-rich manipulation tasks and achieves over 50\% performance improvement compared to state-of-the-art visuomotor policy methods. For result videos, see https://adaptive-compliance.github.io/

Paper Structure

This paper contains 13 sections, 1 theorem, 7 equations, 8 figures, 2 tables.

Table of Contents

  1. INTRODUCTION
  2. RELATED WORK
  3. Method
  4. Preliminaries: Modeling
  5. Demonstration Collection
  6. Estimating Compliance from Demonstrations
  7. Adaptive Compliance Policy
  8. Inputs and Encoding
  9. Outputs and decoding
  10. Experiments
  11. Task I: Item Flipping
  12. Task II: Vase Wiping
  13. CONCLUSIONS

Key Result

Theorem 1

For a robot under external contact described by Eq. eq: newtons_law, there exists a solution $v$ that satisfies the contact constraint eq:contact_constraints as long as it does not control its velocity in the direction of feedback force $f$ in the generalized space.

Figures (8)

  • Figure 1: Compliance Requirements. [Left] Flipping an item requires the robot to follow an arc trajectory (blue) while maintaining contact force. This demands low stiffness in pushing directions ($K_2$) and high stiffness elsewhere ($K_1$). [Right] Wiping a vase necessitates 3D compliance adjustments in both end-effectors to 1) hold the vase, 2) trace the marking, and 3) apply appropriate force without damage. Our algorithm aims to model these spatial-, temporal-, and task-dependent compliance requirements from human demonstration data.
  • Figure 2: Method Comparisons. [LEFT] shows the comparison between a) a typical visuomotor policy chi2023diffusionpolicy, b) a typical force-based compliant policy lee2019making, and c) Adaptive Compliance Policy. [Right] Visualization of virtual target (orange sqaures) and reference poses (yellow circles) inferred by Adaptive Compliance Policy. The directional difference (orange arrows) between the virtual and reference poses encodes compliance direction.
  • Figure 3: Data Collection with Haptic Feedback. We designed a kinesthetic teaching system with low-stiffness compliance that allows the operator to demonstrate variable compliance behavior with direct haptic feedback.
  • Figure 4: Pinching Examples. Grey shape represents a robot tool, blue shape represents a frictionless environment. First three examples are not pinching contact, the last one is.
  • Figure 5: Flipping Scenarios. We test the policy under a variety of settings that require the policy to adapt to different and unseen object geometries (a,b), configuration (c,d) and react to unexpected perturbations caused by fixture movements (e).
  • ...and 3 more figures

Theorems & Definitions (1)

  • Theorem 1