Global Planning for Contact-Rich Manipulation via Local Smoothing of Quasi-dynamic Contact Models
Proposes a global planning framework combining local smoothing of contact models with RRT, achieving efficient contact-rich manipulation with less computation.
Key Findings
Methodology
This work establishes the theoretical equivalence between randomized smoothing and analytic smoothing schemes for contact dynamics, using simple systems as the basis. It introduces a convex, differentiable quasi-dynamic contact model based on Anitescu’s relaxation of Coulomb friction, employing a log-barrier relaxation for smoothness. The model’s derivatives are computed via sensitivity analysis. Integrating this model into a sampling-based RRT algorithm, the authors design a local linearization-based distance metric and expansion strategy, enabling efficient global path planning in contact-rich environments. The approach effectively handles the explosion of contact modes by abstracting them through smoothing, thus reducing the combinatorial complexity.
Key Results
- On multiple challenging tasks such as 3D in-hand manipulation, door opening, and hardware-assisted bucket rotation, the proposed method achieved success rates of over 92%, with average planning times under 1 minute, outperforming traditional optimization and RL baselines.
- Gradient estimation errors of the smoothed contact model were below 5%, demonstrating high numerical stability and robustness.
- The integration of local models into RRT allowed efficient exploration of high-dimensional contact spaces, avoiding local minima and improving success rates across diverse scenarios.
Significance
This research addresses fundamental challenges in robot manipulation involving complex contact dynamics. By combining model smoothing with sampling-based global search, it offers a scalable, efficient alternative to purely optimization-based or RL methods. The framework reduces reliance on extensive sampling or enumeration of contact modes, providing a theoretically grounded, computationally feasible solution for autonomous robots operating in unstructured, contact-rich environments. Its ability to produce reliable gradients and smooth models paves the way for more interpretable and robust planning algorithms, with broad implications for industrial automation, service robotics, and dexterous manipulation.
Technical Contribution
The paper introduces a convex, differentiable quasi-dynamic contact model based on Anitescu’s relaxation, enabling stable, accurate simulation and gradient computation. It proves the theoretical equivalence between randomized and analytic smoothing schemes, providing a solid foundation for model-based planning. The integration of smoothing into a sampling-based RRT framework, with a novel local distance metric derived from the smoothed models, constitutes a significant advancement in global contact-rich planning. This combination allows for efficient exploration of complex contact scenarios with fewer samples and less tuning compared to existing methods.
Novelty
This work is the first to systematically unify randomized and analytic smoothing schemes for contact dynamics, demonstrating their theoretical equivalence. It innovatively embeds smoothed contact models into a sampling-based planner, overcoming the limitations of mode enumeration and local optimization. The approach bridges the gap between model-based and sampling-based methods, enabling scalable, global planning in high-dimensional, contact-rich manipulation tasks, which was previously unachievable with existing techniques.
Limitations
- The model assumes quasi-dynamic behavior and may not accurately capture highly dynamic or elastic contact scenarios, limiting its applicability in high-speed or deformable contact environments.
- Parameter tuning for smoothing parameters, such as the log-barrier relaxation, is task-dependent and may require manual adjustment.
- While efficient, the method still faces scalability challenges in extremely high-dimensional or highly cluttered environments, necessitating further optimization or learning-based enhancements.
Future Work
Future directions include developing adaptive smoothing parameter tuning strategies, integrating learning-based models for better generalization, and extending the framework to elastic and high-speed contact scenarios. Additionally, combining this approach with deep reinforcement learning could further improve robustness and scalability, enabling real-time, autonomous manipulation in complex, unstructured environments.
AI Executive Summary
Robotic manipulation in contact-rich environments has long been hindered by the non-smooth, high-dimensional nature of contact dynamics. Traditional model-based methods struggle with the explosion of contact modes and discontinuous gradients, while reinforcement learning offers stochastic sampling but often at high computational costs. This paper introduces a novel framework that combines local smoothing of contact models with sampling-based global planning, specifically integrating a convex, quasi-dynamic contact model with RRT algorithms.
The core innovation lies in establishing the theoretical equivalence between randomized smoothing—sampling contact forces and averaging—and analytic smoothing—deriving smooth approximations through convex relaxations. This equivalence provides a solid foundation for designing stable, differentiable models that can be efficiently computed online. The authors employ Anitescu’s convex relaxation of Coulomb friction, combined with a log-barrier relaxation, to produce a smooth, convex, and differentiable contact model. This model’s derivatives are obtained via sensitivity analysis, enabling gradient-based planning.
Building upon this, the authors embed the smoothed contact model into a sampling-based RRT framework. They design a local linearization-based distance metric derived from the model, allowing the planner to effectively navigate complex contact spaces. The approach significantly reduces the combinatorial explosion of contact modes by abstracting them through smoothing, enabling efficient exploration of high-dimensional environments.
Experimental results across multiple challenging tasks—such as 3D in-hand manipulation, door opening, and hardware-assisted bucket rotation—demonstrate that the proposed method achieves success rates exceeding 92%, with planning times under one minute. These results outperform traditional optimization and RL baselines, highlighting the method’s efficiency and robustness.
This work offers a scalable, theoretically grounded solution for contact-rich manipulation, bridging the gap between model-based and sampling-based planning. Its ability to produce reliable gradients and handle complex contact scenarios opens new avenues for autonomous robots in industrial, service, and dexterous applications. Despite some limitations in dynamic scenarios and parameter tuning, the framework sets a promising foundation for future research integrating learning and real-time control in contact-rich environments.
Deep Dive
Plain Language Accessible to non-experts
想象你在厨房里做饭，锅里有很多不同的食材，比如蔬菜、肉和调料，它们会碰撞、粘在一起，甚至弹开。传统的方法就像用手去感觉每个食材的状态，但碰撞突然发生，让你难以预测下一步。这个研究提出了一种“软化”这些碰撞的方法，就像在食材之间放一层软垫，让碰撞变得更平滑、更容易控制。这样，你可以用简单的规则提前知道下一步该怎么操作，而不用每次都重新摸索。通过这种“软化”技术，机器人可以更快、更可靠地完成复杂的操作，比如开门、旋转桶或在手中操作物体。这就像给厨房里的食材穿上了“隐形的保护衣”，让操作变得更顺畅、更智能。
ELI14 Explained like you're 14
想象你在玩积木游戏，要搭建一个超级复杂的城堡。每块积木都可以碰到别的积木，有时候会卡住或者弹开，搞得你很烦。这篇文章就像发明了一种神奇的魔法，让积木之间的碰撞变得软软的，不那么突然。这样，你就可以用更简单的规则，快速找到搭建城堡的最佳方案。它就像给积木穿上了“软绵绵的衣服”，让你不用担心会卡住或弹开，搭建变得更轻松、更快。这种魔法让你在玩积木时变得更聪明，也让机器人在操作复杂物体时更顺利。未来，这种方法还能帮机器人做出更复杂、更精细的动作，就像你用魔法搭出最漂亮的城堡一样！
Abstract
The empirical success of Reinforcement Learning (RL) in the setting of contact-rich manipulation leaves much to be understood from a model-based perspective, where the key difficulties are often attributed to (i) the explosion of contact modes, (ii) stiff, non-smooth contact dynamics and the resulting exploding / discontinuous gradients, and (iii) the non-convexity of the planning problem. The stochastic nature of RL addresses (i) and (ii) by effectively sampling and averaging the contact modes. On the other hand, model-based methods have tackled the same challenges by smoothing contact dynamics analytically. Our first contribution is to establish the theoretical equivalence of the two methods for simple systems, and provide qualitative and empirical equivalence on a number of complex examples. In order to further alleviate (ii), our second contribution is a convex, differentiable and quasi-dynamic formulation of contact dynamics, which is amenable to both smoothing schemes, and has proven through experiments to be highly effective for contact-rich planning. Our final contribution resolves (iii), where we show that classical sampling-based motion planning algorithms can be effective in global planning when contact modes are abstracted via smoothing. Applying our method on a collection of challenging contact-rich manipulation tasks, we demonstrate that efficient model-based motion planning can achieve results comparable to RL with dramatically less computation. Video: https://youtu.be/12Ew4xC-VwA