ComFree-Sim: Multi-Domain Simulation Framework

• ComFree-Sim is a label family representing distinct systems: a GPU-parallel analytical contact physics engine for robotics, an end-to-end cell-free 6G simulator, and SIM-assisted MIMO architectures.
• It leverages closed-form dual-cone updates in robotics, interactive ray casting in network simulations, and layered wave-domain beamforming to achieve enhanced computational efficiency and scalability.
• Each implementation faces domain-specific trade-offs, including approximations in physical modeling and limitations in wireless channel analysis under varied operational conditions.

ComFree-Sim is a non-unique designation that appears in recent arXiv literature for multiple technically distinct systems. The name most explicitly refers to a GPU-parallelized analytical contact physics engine for contact-rich robotics (Borse et al., 12 Mar 2026), but it is also used for an open-source end-to-end simulator for practical cell-free 6G network deployments, presented as LuSim (Tärneberg et al., 2023), and for several SIM-enabled cell-free massive MIMO / holographic MIMO architectures in which stacked intelligent metasurfaces (SIMs) perform wave-domain beamforming at distributed access points (Park et al., 23 Jun 2025). This suggests that “ComFree-Sim” is best treated as a label family whose meaning depends on disciplinary context.

1. Terminological scope

The literature uses the label across at least three domains.

Usage	Paper	Core description
Contact-rich robotics	(Borse et al., 12 Mar 2026)	GPU-parallel analytical contact physics engine built on complementarity-free contact modeling
Cell-free 6G network simulation	(Tärneberg et al., 2023)	Open-source end-to-end simulation platform, presented as LuSim
SIM-assisted cell-free MIMO/HMIMO	(Park et al., 23 Jun 2025, Li et al., 2024, Shi et al., 2024)	Cell-free architectures using stacked intelligent metasurfaces for wave-domain beamforming

A recurrent source of ambiguity is the acronym SIM. In the wireless cell-free papers, SIM denotes stacked intelligent metasurfaces (Park et al., 23 Jun 2025). In adjacent cellular-systems work, by contrast, SIM denotes the Subscriber Identity Module, as in SIM tracing and SIM tunneling platforms (Gegenhuber et al., 25 Jun 2025). The same surface form therefore spans robotics simulation, radio-network simulation, metasurface-assisted communication architectures, and SIM-card instrumentation.

2. ComFree-Sim as a contact-rich robotics physics engine

In robotics, ComFree-Sim is a GPU-parallel analytical contact physics engine designed to remove the main bottleneck in contact-rich simulation: iterative complementarity-based contact resolution (Borse et al., 12 Mar 2026). The core idea is to avoid solving a nonlinear complementarity problem or constrained optimization problem at every timestep and instead compute contact impulses in closed form using a complementarity-free, impedance-style prediction–correction update in the dual cone of Coulomb friction.

The discrete-time dynamics are written as

$$\boldsymbol{M}(\boldsymbol{q})\,d\boldsymbol{v} + \boldsymbol{c}(\boldsymbol{q},\boldsymbol{v})\,dt = \boldsymbol{\tau}\,dt + \mathbf{J}(\boldsymbol{q})^{\top}\boldsymbol{\lambda}\,dt ,$$

with smooth prediction

$$\boldsymbol{v}_{smooth}^+ := \boldsymbol{v} + \boldsymbol{M}^{-1}(\boldsymbol{\tau}-\boldsymbol{c})\,dt.$$

Contact correction is then imposed through dual-cone inequalities in velocity space rather than through force-space complementarity. The formulation is extended to a unified 6D contact model covering tangential friction, torsional friction, and rolling friction, with primal constraints

$$\|\boldsymbol{\lambda}^{\mathrm{t}}\| \le \mu^{\mathrm{t}}\lambda^{n},\quad |m^{\mathrm{tor}}| \le \mu^{\mathrm{tor}}\lambda^{n},\quad \|\boldsymbol{m}^{\mathrm{roll}}\| \le \mu^{\mathrm{roll}}\lambda^{n}.$$

To make the update computationally practical, each quadratic dual cone is approximated by a polyhedral cone, after which contact impulses are obtained analytically by an elementwise ReLU-type activation on violating faces.

The engine’s computational advantage follows from separability. Contact computation is separable across contact pairs and separable across cone facets, which maps naturally to GPU kernels. The implementation uses Warp and exposes a MuJoCo-compatible interface as a drop-in backend alternative to MuJoCo Warp (MJWarp). The paper reports near-linear scaling with contact count, 2–3 times higher throughput in dense contact scenes, and about 3× faster simulation speed in dense-contact scaling tests. In a penetration benchmark, reported mean penetration depth was $1.7 \pm 4.9$ mm for MJWarp and as low as $0.9 \pm 1.5$ mm for ComFree-Sim at tuned settings. In real-time MPPI-based MPC on a physical LEAP Hand, ComFree-Sim reduced MPPI compute time by about 2.4× on average, improved closed-loop success rate by about 27 percentage points on average, and enabled 35–72 Hz control on hardware. The paper also reports dynamics-aware motion retargeting results using SPIDER on five Unitree G1 motions, with task-level performance comparable to MJWarp but faster optimization times.

3. ComFree-Sim as an end-to-end simulator for cell-free 6G deployments

In wireless-network simulation, ComFree-Sim is presented as an open-source simulator called LuSim for practical cell-free 6G network deployments (Tärneberg et al., 2023). Its purpose is to bridge a gap between realistic physical propagation and system-level network behavior in a single framework. The architecture consists of a Unity-based ray-casting physical-layer engine and a Python-based system simulator. Configuration is externalized through JSON/YAML-style scenario definitions, and the two halves communicate through a ZeroMQ proxy over UDP. On the system side, the simulator uses SimPy inside the RW infrastructure for high-level discrete-event simulation.

Unity is not used merely for rendering. It acts as the physical-layer environment in which buildings, surfaces, users, and base stations are placed and from which rays are cast. The paper emphasizes interactive ray casting, interactive visualization, and GPU acceleration. The simulator can run either as a standalone Unity application or within the Unity editor, which supports repeatable experiments as well as interactive debugging and visualization.

The channel-generation workflow is built around a geometry-based stochastic channel model (GSCM) initialized in Unity through multi-path components (MPCs). The described steps are: MPC distribution, MPC filtering, surface association, path-gain modeling, parameter estimation, look-up table generation, and scene visualization. MPCs are uniformly distributed over the environment, especially along building facades, for the first, second, and third reflection orders; densities are chosen according to a COST-based model. The resulting model captures distance-dependent path loss as a classical log-distance power law, Gamma-distributed shadow fading, time correlation of fading through an exponential autocorrelation structure, angular dependence of scattering, obstruction and diffraction around corners, penetration losses, dynamic LOS determination for moving users, and multi-order reflections.

The system-level simulator consumes channel realizations, DSS-compliant channel measurements, and entity positions to evaluate radio resource allocation, centralized optimization, dynamic federation formation, energy-aware antenna disabling, wireless power transfer scheduling, and localization / sensing improvements. The paper states that the simulator can represent digital twins of the network infrastructure including energy, latency and other models. Supported scenarios include urban outdoor environments, indoor 3D settings, fully 3D cell-free deployments, moving users, and large antenna counts with distributed antenna layouts. Validation is reported against real-world measurement data for channel gain, Doppler spread, and delay spread, including urban intersection scenarios not used during parameter estimation.

4. ComFree-Sim as SIM-enabled cell-free MIMO and HMIMO

A third usage refers to cell-free architectures in which stacked intelligent metasurfaces are deployed at access points to shift part of beamforming from the digital domain to the electromagnetic domain. In the fronthaul-constrained CF-mMIMO formulation, each access point is equipped with a SIM consisting of $L$ metasurface layers with $M$ meta-atoms per layer, and the end-to-end SIM transformation is written as

$$\mathbf{G}_i = \boldsymbol{\Phi}_{i,1}\mathbf{W}_{i,2}\boldsymbol{\Phi}_{i,2}\cdots \mathbf{W}_{i,L}\boldsymbol{\Phi}_{i,L}.$$

The paper formulates joint optimization of digital beamforming, wave-domain beamforming, and fronthaul compression under finite-capacity fronthaul constraints for both uplink and downlink. Because the resulting problems are high-dimensional and non-convex, the solution uses alternating optimization, with digital blocks handled through the matrix Lagrangian duality transform and Fenchel’s inequality, and wave-domain blocks optimized either layer-by-layer or by gradient ascent (Park et al., 23 Jun 2025). Numerical results report fast convergence, near fully-digital performance with sufficiently deep SIMs, especially around $L=7$ layers, and a runtime reduction of more than a factor of 20 in the simulations. The same study notes that the downlink gap to fully digital is larger at low SNR.

In the uplink cell-free HMIMO formulation, each AP has $M$ receive antennas and a SIM with $$\boldsymbol{v}_{smooth}^+ := \boldsymbol{v} + \boldsymbol{M}^{-1}(\boldsymbol{\tau}-\boldsymbol{c})\,dt.$$0 layers, each containing $$\boldsymbol{v}_{smooth}^+ := \boldsymbol{v} + \boldsymbol{M}^{-1}(\boldsymbol{\tau}-\boldsymbol{c})\,dt.$$1 passive reconfigurable elements. The cascaded effective channel is

$$\boldsymbol{v}_{smooth}^+ := \boldsymbol{v} + \boldsymbol{M}^{-1}(\boldsymbol{\tau}-\boldsymbol{c})\,dt.$$2

and APs optimize local SIM coefficients and local receiver combiners using only local CSI, while the CPU fuses local detections using an MMSE criterion (Li et al., 2024). The paper explicitly models RF-chain hardware impairments at both UEs and APs and shows that they limit achievable rate in the high-SNR regime. Its reported setup uses $$\boldsymbol{v}_{smooth}^+ := \boldsymbol{v} + \boldsymbol{M}^{-1}(\boldsymbol{\tau}-\boldsymbol{c})\,dt.$$3 GHz, $$\boldsymbol{v}_{smooth}^+ := \boldsymbol{v} + \boldsymbol{M}^{-1}(\boldsymbol{\tau}-\boldsymbol{c})\,dt.$$4 APs, $$\boldsymbol{v}_{smooth}^+ := \boldsymbol{v} + \boldsymbol{M}^{-1}(\boldsymbol{\tau}-\boldsymbol{c})\,dt.$$5 UEs, $$\boldsymbol{v}_{smooth}^+ := \boldsymbol{v} + \boldsymbol{M}^{-1}(\boldsymbol{\tau}-\boldsymbol{c})\,dt.$$6 receive antennas, $$\boldsymbol{v}_{smooth}^+ := \boldsymbol{v} + \boldsymbol{M}^{-1}(\boldsymbol{\tau}-\boldsymbol{c})\,dt.$$7 elements per layer, and $$\boldsymbol{v}_{smooth}^+ := \boldsymbol{v} + \boldsymbol{M}^{-1}(\boldsymbol{\tau}-\boldsymbol{c})\,dt.$$8 layers. It reports that the layer-by-layer iterative algorithm converges within roughly 10 iterations, and that about 4-bit phase resolution can approach infinite-resolution performance.

In a related low-power and cost formulation, the SIM-enhanced CF mMIMO system uses TDD, a two-layer signal processing framework, phase-aware MMSE channel estimation, MR combining at APs, and LSFD or EGCD at the CPU (Shi et al., 2024). The SIM at AP $$\boldsymbol{v}_{smooth}^+ := \boldsymbol{v} + \boldsymbol{M}^{-1}(\boldsymbol{\tau}-\boldsymbol{c})\,dt.$$9 applies diagonal phase-shift matrices

$$\|\boldsymbol{\lambda}^{\mathrm{t}}\| \le \mu^{\mathrm{t}}\lambda^{n},\quad |m^{\mathrm{tor}}| \le \mu^{\mathrm{tor}}\lambda^{n},\quad \|\boldsymbol{m}^{\mathrm{roll}}\| \le \mu^{\mathrm{roll}}\lambda^{n}.$$0

and the overall SIM beamforming matrix is

$$\|\boldsymbol{\lambda}^{\mathrm{t}}\| \le \mu^{\mathrm{t}}\lambda^{n},\quad |m^{\mathrm{tor}}| \le \mu^{\mathrm{tor}}\lambda^{n},\quad \|\boldsymbol{m}^{\mathrm{roll}}\| \le \mu^{\mathrm{roll}}\lambda^{n}.$$1

The paper proposes an interference-based greedy pilot assignment, wave-based beamforming using only statistical CSI, and max-min SE power control solved by bisection. Reported numerical findings include a 57\% SE improvement for the proposed wave-based beamforming algorithm, best SE performance with 20 APs plus 1200 SIM meta-atoms, the claim that 10 APs with SIM can outperform 15 APs in a traditional CF mMIMO system, and that a 2-antenna SIM-enhanced AP can achieve performance close to a traditional 4-antenna AP. The same paper reports the best meta-atom spacing as

$$\|\boldsymbol{\lambda}^{\mathrm{t}}\| \le \mu^{\mathrm{t}}\lambda^{n},\quad |m^{\mathrm{tor}}| \le \mu^{\mathrm{tor}}\lambda^{n},\quad \|\boldsymbol{m}^{\mathrm{roll}}\| \le \mu^{\mathrm{roll}}\lambda^{n}.$$2

5. Methodological patterns and reported limitations

Across these usages, a common design pattern is the replacement of expensive canonical formulations by structured surrogates or decompositions. In robotics, iterative complementarity solves are replaced by a closed-form dual-cone impedance update (Borse et al., 12 Mar 2026). In LuSim, exhaustive ray tracing is replaced by interactive ray casting combined with a GSCM (Tärneberg et al., 2023). In SIM-enabled cell-free MIMO, one RF chain per physical antenna is avoided by shifting part of beamforming to the wave domain through stacked intelligent metasurfaces (Park et al., 23 Jun 2025).

The limitations are correspondingly domain-specific. LuSim is described as a pragmatic middle ground: ray casting and GSCM are more efficient than exhaustive ray tracing, but still approximate the full electromagnetic complexity of real-world propagation, and the discussion identifies RIS and integrated sensing and communications as future extensions rather than fully mature features (Tärneberg et al., 2023). The SIM-assisted HMIMO formulation assumes narrowband transmission, independent SIM layers without mutual coupling, local CSI at each AP, and idealized controllable diagonal SIM layers; it also states that wideband/spatial-wideband effects are not addressed and that HWIs and layer attenuation limit performance at high SNR (Li et al., 2024). The fronthaul-constrained hybrid digital-wave formulation reports that the gap to fully digital remains more pronounced in the downlink at low SNR (Park et al., 23 Jun 2025). For robotics, ComFree-Sim is reported to tolerate moderate timestep sizes, but it generally benefits from smaller $$\|\boldsymbol{\lambda}^{\mathrm{t}}\| \le \mu^{\mathrm{t}}\lambda^{n},\quad |m^{\mathrm{tor}}| \le \mu^{\mathrm{tor}}\lambda^{n},\quad \|\boldsymbol{m}^{\mathrm{roll}}\| \le \mu^{\mathrm{roll}}\lambda^{n}.$$3 than MJWarp (Borse et al., 12 Mar 2026).

A related misconception is to treat all occurrences of “ComFree-Sim” as references to one simulator. The cited record shows instead that the same label is attached to a robotics backend, a cell-free 6G end-to-end simulator, and several metasurface-assisted cell-free network architectures.

6. Related systems and adjacent terminology

Two adjacent systems help situate the term. SIMulator is a low-cost SIM tracing platform that reproduces the essential functions of traditional SIM tracing hardware using simple components such as UART interfaces, GPIO ports, and a Raspberry Pi Pico (~4 USD) (Gegenhuber et al., 25 Jun 2025). Its key architectural choice is to electrically decouple the SIM side from the modem side and bridge them only at the APDU level. The path is Modem ↔ Pico ↔ Host relay ↔ SIM provider ↔ physical SIM/eSIM, with support for physical SIM readers, Android SIM Access Profile (SAP), T=0 and T=1 contact smart cards, and ISO 7816 Waiting Time eXtensions (WTX). The paper reports successful testing with 7 modems, 4 smart card terminals, and 1,000 ms artificial delay without failures or degraded behavior. It is explicitly described as not a direct implementation of ComFree-Sim, but as a related enabling system and a key building block for low-cost SIM tunnel / SIM virtualization ideas.

SimRIS Channel Simulator is an open-source MATLAB package with GUI for mmWave RIS-assisted communication systems (Basar et al., 2020). It models the cascaded RIS link and direct path through

$$\|\boldsymbol{\lambda}^{\mathrm{t}}\| \le \mu^{\mathrm{t}}\lambda^{n},\quad |m^{\mathrm{tor}}| \le \mu^{\mathrm{tor}}\lambda^{n},\quad \|\boldsymbol{m}^{\mathrm{roll}}\| \le \mu^{\mathrm{roll}}\lambda^{n}.$$4

supports indoor and outdoor scenarios, and is demonstrated at 28 GHz with support also stated for 73 GHz. Its channel model includes RIS element radiation characteristics, path loss, LOS probability, and a 3GPP-style clustered mmWave model. Although it is not a cell-free simulator, it supplies a relevant channel-modeling baseline for RIS-oriented and metasurface-oriented wireless studies.

Taken together, these neighboring systems clarify the semantic range around the term. In one branch, ComFree-Sim concerns analytical contact physics; in another, it denotes cross-layer cell-free 6G simulation; in another, it denotes SIM-assisted cell-free radio architectures; and in adjacent cellular-instrumentation work, “SIM” refers not to metasurfaces but to the Subscriber Identity Module. Accurate interpretation therefore depends on the paper’s disciplinary setting, the expansion of “SIM,” and whether the object under discussion is a simulator, a communication architecture, or an enabling hardware platform.