Abstract
This paper presents a novel simulation platform, ZeMa, designed for robotic manipulation tasks concerning soft objects. Such simulation ideally requires three properties: two-way soft-rigid coupling, intersection-free guarantees, and frictional contact modeling, with acceptable runtime suitable for deep and reinforcement learning tasks. Current simulators often satisfy only a subset of these needs, primarily focusing on distinct rigid-rigid or soft-soft interactions. The proposed ZeMa prioritizes physical accuracy and integrates the incremental potential contact method, offering unified dynamics simulation for both soft and rigid objects. It efficiently manages soft-rigid contact, operating 75x faster than baseline tools with similar methodologies like IPC-GraspSim. To demonstrate its applicability, we employ it for parallel grasp generation, penetrated grasp repair, and reinforcement learning for grasping, successfully transferring the trained RL policy to real-world scenarios.
Introduction
In robot simulation, maintaining high frame rates often comes at the cost of physical accuracy. Popular engines like Bullet, MuJoCo, or Isaac Gym frequently allow objects to slightly interpenetrate during collisions. While this "ghosting" is acceptable for rigid-body games, it is problematic for simulating soft objects or precision grasping, where the exact contact surface determines success or failure.
Existing solutions are typically either fast but physically inaccurate, or accurate but extremely slow. Furthermore, most simulators use separate solvers for rigid and soft bodies, struggling with complex frictional interactions. ZeMa addresses these challenges by providing two-way soft-rigid coupling, guaranteed zero interpenetration, and realistic friction, all at a speed suitable for AI training.
Methodology: Unified Incremental Potential Contact
The core of ZeMa is a unified energy formulation based on the Incremental Potential Contact (IPC) method.
Unified Representation
Instead of treating rigid and soft bodies as distinct entities, ZeMa treats them as part of a single optimization problem. Soft bodies are modeled as tetrahedral meshes (FEM), while rigid bodies are represented as affine bodies—objects with such high stiffness that they do not deform, yet they utilize the same solver framework as soft bodies. This ensures that the two-way coupling is mathematically seamless.
Intersection-Free Guarantees
To prevent intersections, ZeMa uses a barrier-augmented potential energy function. As two objects approach each other, the barrier energy grows to infinity, making interpenetration mathematically impossible. This is critical for high-fidelity grasping where the contact manifold must be precisely modeled.
Frictional Contact and Acceleration
ZeMa employs a variational approximation of the Coulomb friction model, integrating friction directly into the optimization framework. To overcome the traditional slowness of IPC, the authors implemented a GPU-parallelized interface, allowing for simultaneous handling of multiple grasp attempts or reinforcement learning (RL) environments.
Results and Applications
Performance and Accuracy
ZeMa is 8x faster than IPC-GraspSim in single-threaded scenarios and over 75x faster with its GPU-parallelized interface. In sim-to-real experiments using a gel elastomer, ZeMa’s predicted deformations matched real-world outcomes with a relative error of only 5.1% to 5.5%.
Grasp Repair
A significant application of ZeMa is "healing" existing datasets. Many "successful" grasps in datasets like DexGraspNet actually contain physically impossible overlaps. ZeMa can shift these grasps into the nearest valid, intersection-free state, reducing average penetration depth from 3.69 cm to 0 cm.
Reinforcement Learning
A Franka Emika Panda robot was trained in ZeMa to pick up cubes. The policy achieved a 73% success rate in simulation and transferred to a 70% success rate on a real robot without further tuning, demonstrating the high fidelity and transferability of the simulation.
Conclusion
ZeMa represents a significant advancement in high-fidelity soft-robotics simulation. By leveraging IPC and GPU acceleration, it makes intersection-free manipulation viable for large-scale data generation and reinforcement learning. Future work will focus on optimizing Discrete Collision Detection (DCD) and linear system solvers to further bridge the gap toward real-time performance.
Summary updated