ComFree-Sim: A GPU-Parallelized Analytical Contact Physics Engine for Scalable Contact-Rich Robotics Simulation and Control

In the field of robotics, physics simulation is a critical research direction. Early physics engines like ODE and Bullet were primarily developed for animation and games, focusing on visually plausible motion rather than high-fidelity contact dynamics. As the complexity of robotic tasks increased, robotics-oriented simulators like MuJoCo and Drake emerged, emphasizing efficiency and physics fidelity for planning, control, and policy learning. In particular, MuJoCo enforces full friction-cone constraints and casts contact resolution as a cone complementarity problem solvable via convex optimization. However, driven by large-scale training and hardware acceleration needs, these simulators have been further ported to GPUs/TPUs to improve parallel simulation throughput.

In contact-rich robotic tasks, the efficiency and accuracy of contact resolution are core issues. Traditional contact resolution methods rely on complementarity constraints or constrained optimization, whose computational cost grows superlinearly with contact density, limiting their feasibility in real-time applications. Even state-of-the-art GPU-parallelized physics engines like MJWarp exhibit substantial step-time growth in dense contact scenes, complicating real-time deployment.

The core innovations of ComFree-Sim include its complementarity-free contact modeling approach. First, this method calculates contact impulses in closed form via an impedance-based prediction-correction update in the dual cone of Coulomb friction. Second, contact computation is decoupled across contact pairs and separated across cone facets, naturally mapping to GPU kernels, achieving near-linear runtime scaling with the number of contacts. Finally, ComFree-Sim extends to a unified 6D contact model capturing tangential, torsional, and rolling friction, introducing a practical dual-cone impedance heuristic.

The methodology of ComFree-Sim includes the following key steps:

• �� Employ a complementarity-free contact modeling approach, calculating contact impulses in closed form via an impedance-based prediction-correction update in the dual cone of Coulomb friction.
• �� Decouple contact computation across contact pairs and separate it across cone facets, naturally mapping to GPU kernels, achieving near-linear runtime scaling with the number of contacts.
• �� Extend to a unified 6D contact model capturing tangential, torsional, and rolling friction, introducing a practical dual-cone impedance heuristic.
• �� Implement ComFree-Sim in the Warp programming framework and expose it through a MuJoCo-compatible interface as a drop-in backend alternative to MJWarp.

The experimental design includes benchmarking ComFree-Sim against MJWarp in dense contact scenarios, evaluating penetration depth, friction behaviors, stability, and simulation runtime scaling. All tests run on an AMD 32-core CPU with an NVIDIA RTX 4090 GPU. Experiments also include real-time MPC on a real-world multi-fingered LEAP hand to assess ComFree-Sim's performance in practical contact-rich tasks.

Experimental results show that ComFree-Sim demonstrates outstanding performance in dense contact scenarios, achieving 2-3 times higher throughput than MJWarp, with comparable physical fidelity. ComFree-Sim outperforms or matches MJWarp in penetration depth, friction behaviors, stability, and simulation runtime scaling, especially in contact-rich scenarios. In real-time MPC experiments on a real-world multi-fingered LEAP hand, ComFree-Sim achieves higher closed-loop success rates and enables practical high-frequency control.

ComFree-Sim is suitable for robotic tasks requiring high-frequency control and real-time response, such as dexterous manipulation and motion planning. Its GPU acceleration and complementarity-free modeling significantly enhance simulation efficiency while maintaining physical realism, particularly in dense contact scenarios. This breakthrough has important implications for both academia and industry, especially in real-time control and large-scale data generation.

Despite ComFree-Sim's outstanding performance in many aspects, it may still encounter performance bottlenecks in extremely high contact density scenarios, especially when GPU resources are limited. Additionally, the method may have limitations in accurately simulating certain complex friction behaviors, requiring further experimental validation. In some specific robotic applications, targeted parameter adjustments may be needed to achieve optimal performance. Future research directions include further optimizing the GPU implementation of ComFree-Sim to support larger-scale simulation scenarios.