Switch: Routing, Gating, and Control in Modern Systems

• Switch is a mechanism that selects among alternative ports, states, or flows in both physical and algorithmic environments, extending beyond a simple binary contact.
• Recent research presents diverse switching architectures including integrated photonics, continuous-variable quantum networks, and cryogenic MEMS controllers, each quantified by metrics like extinction ratio, crosstalk, and speed.
• Switching frameworks are modeled as constrained selection problems under domain-specific limits, influencing design choices in optical routing, entanglement swapping, and robotic skill transitions.

Searching arXiv for the cited switch-related papers and recent context. In recent arXiv literature, a switch denotes a mechanism, device, architecture, or control framework that selects among alternative ports, states, communication flows, or behaviors. The term spans integrated photonic routers, fiber-based single-photon interferometric routers, continuous-variable quantum repeater nodes that schedule entanglement flows, room-temperature controllers for cryogenic MEMS RF networks, hierarchical humanoid systems for online skill transitions, and embodied-AI benchmarks centered on tangible control interfaces such as light switches and appliance panels (Ye et al., 2015, Stern et al., 2015, Alarcón et al., 2020, Tillman et al., 2022, Spietz et al., 6 Jan 2025, Lau et al., 16 Apr 2026, Lin et al., 20 Nov 2025). Taken together, these works indicate that a switch is not restricted to a binary electrical contact: it is a broader operator for routing, gating, scheduling, or verifying change in a physical or algorithmic system.

1. Functional scope and core abstraction

Across the cited works, switching has a common operational form: a system receives an input state, command, or demand, and then causes one among several admissible outcomes. In integrated optics, the outcome is a BAR or CROSS path in a waveguide network, or one output among many in a switch fabric (Ye et al., 2015, Sun et al., 16 Feb 2025). In a fiber-optical Sagnac interferometer, the outcome is a single photon directed to detector $D_1$ or $D_2$ by controlling an internal phase difference (Alarcón et al., 2020). In continuous-variable quantum networking, the outcome is a selected bipartite entanglement flow served by a repeater node under a Max-Weight scheduling policy (Tillman et al., 2022). In cryogenic RF control, the outcome is a selected DSUB pin driven by a $90$ V line through a relay matrix (Spietz et al., 6 Jan 2025). In humanoid control, the outcome is a feasible path through a multi-skill state graph that realizes a commanded skill transition (Lau et al., 16 Apr 2026). In embodied evaluation, the outcome is a correct prediction or action over a tangible control interface, including state recognition, action generation, state-transition prediction, and result verification (Lin et al., 20 Nov 2025).

This suggests a unifying abstraction: switching is a constrained selection problem under physical, informational, or safety limits. The constraints differ by domain—optical loss, crosstalk, queue stability, relay actuation, dynamic feasibility, or partial observability—but the underlying role remains the same.

2. Photonic and optical switching architectures

Integrated photonic switching is represented in the cited literature by three distinct design lineages: compact plasmonic $2\times2$ electro-optic switching, mode- and wavelength-aware silicon routing, and larger switch-and-select fabrics in a tri-layer platform. A related quantum-optical lineage appears in the fiber-based single-photon Sagnac switch.

The plasmonic MOS-based $2\times2$ electro-optic switch is a three-waveguide directional-coupler layout with two passive Silicon-on-Insulator waveguide busses and a central “island” waveguide containing an N-doped Si core, a thin Indium Tin Oxide active layer, a SiO$_2$ gate dielectric, and an Al top contact. Its switching mechanism is capacitor-like: under bias, carrier accumulation in the ITO shifts the plasma frequency, changes the complex refractive index, and modifies the supermode structure of the coupler. Without bias, the coupling length is chosen so that light from the BAR input fully transfers to the CROSS output; under bias, the island becomes partially metallic and reflective, and most TM light remains in the BAR waveguide. Reported performance includes extinction ratio of approximately $18$ dB for the CROSS-state and $7$ dB for the BAR-state, insertion loss of approximately $1.3$ dB and $2.4$ dB respectively, energy per bit of approximately $D_2$0–$D_2$1 fJ for $D_2$2–$D_2$3 V, broadband operation across $D_2$4–$D_2$5m, and a device length of approximately $D_2$6m with total area approximately $D_2$7 (Ye et al., 2015). The governing coupler relation is

$D_2$8

The silicon multimode switch for simultaneous mode-division multiplexing and wavelength-division multiplexing uses a “convert-process-reconvert” workflow. Incoming spatial modes are first transformed into the fundamental TE$D_2$9 mode of single-mode waveguides by phase-matched ring resonator couplers; four microring resonators then act as $90$0 switches; after switching, the channels are reconverted into their original spatial modes. The device demonstrates a $90$1 multimode switch that routes four data channels with crosstalk below $90$2 dB, bit-error rates below $90$3, and power penalties below $90$4 dB on all channels at $90$5 Gbps when each channel is input and routed separately. When all four channels are simultaneously routed, the switch exhibits an additional power penalty of less than $90$6 dB (Stern et al., 2015). In this architecture, switching is not merely spatial redirection; it is mode-aware and wavelength-compatible channel processing.

The tri-layer SiN-on-Si $90$7 optical switches implement a spatial-multiplexed switch-and-select topology with a three-dimensional crossing-free shuffle network formed by a tri-layer Si-SiN-SiN structure. The topology is strictly non-blocking and permutation-capable; each path uses exactly two resonant cells, one in the multiplexer and one in the demultiplexer, and first-order crosstalk is suppressed by off-resonance attenuation. The reported $90$8D crossing exhibits only $90$9 dB insertion loss and less than $2\times2$0 dB crosstalk while occupying only $2\times2$1m$2\times2$2m. Two actuator variants are reported: a thermo-optic device with measured on-chip losses from $2\times2$3 to $2\times2$4 dB, $2\times2$5 dB average, and switching time $2\times2$6s; and an electro-optic device with losses between $2\times2$7 and $2\times2$8 dB, $2\times2$9 dB average, and switching time $2\times2$0 ns. The abstract reports crosstalk ranges of $2\times2$1 to $2\times2$2 dB and $2\times2$3 to $2\times2$4 dB for the thermo-optic and electro-optic devices respectively, while the detailed exposition writes these under the convention $2\times2$5 as $2\times2$6 dB to $2\times2$7 dB and $2\times2$8 dB to $2\times2$9 dB (Sun et al., 16 Feb 2025). This sign difference is a notational issue rather than a disagreement about magnitude.

The fiber-optical Sagnac single-photon switch implements polarization-independent routing by placing two fast electro-optical telecom phase modulators inside the loop so that each modulator acts on an orthogonal polarization component. A $_2$0 m fiber spool ensures that only one of the two counter-propagating wavepackets is phase-shifted by the $_2$1 ns electrical pulse. The output probabilities are

$_2$2

which are independent of the polarization amplitudes $_2$3 and $_2$4. The reported average visibility is $_2$5, corresponding to an extinction ratio of approximately $_2$6 dB, in agreement with the reported average extinction ratio of more than $_2$7 dB. The device uses commercial off-the-shelf telecom components, exhibits insertion loss of approximately $_2$8 dB, and employs $_2$9 GHz LiNbO$18$0 phase modulators; the authors drove them with $18$1 ns pulses, while the exposition states that sub-$18$2 ps switching is possible in principle (Alarcón et al., 2020).

3. Continuous-variable quantum switching and entanglement routing

The continuous-variable quantum switch is formulated as a central “hub-and-spoke” repeater node connecting $18$3 end-nodes. Each spoke carries a bipartite entanglement flow request between two end-nodes via the switch. On each spoke, two-mode squeezed vacuum sources generate entangled CV pairs, and the switch and end-nodes share $18$4-fold multiplexed channels per link. Half the channels are oriented with the TMSV source at the end-node and the quantum scissors at the switch, and half vice versa; the exposition states that this symmetry guarantees that, when a herald succeeds, its orientation is equally likely “head-to-tail” for optimized end-to-end rates (Tillman et al., 2022).

Each transmitted mode passes through a lossy channel of transmissivity $18$5, and at the receiving end a quantum-scissor implementation of noiseless linear amplification with gain $18$6 probabilistically heralds an error-corrected link with success probability

$18$7

The overall link-generation success per time step is

$18$8

when $18$9. The switch has a very short memory lifetime of one time-step, so any unused elementary entanglement is discarded at the end of the step, whereas end-nodes have long-lived classical and quantum memories supporting hashing protocols for entanglement distillation over many time-steps. When a flow is selected, the switch performs dual-homodyne Gaussian measurements on the corresponding ports, and entanglement swapping succeeds deterministically in the model, with $7$0 (Tillman et al., 2022).

The scheduling problem is expressed in slotted time with i.i.d. arrivals of rate $7$1 for each flow $7$2, queue lengths $7$3, and link-availability indicators $7$4. A binary vector $7$5 is a matching if each physical port $7$6 is used by at most one flow:

$7$7

The Max-Weight rule chooses

$7$8

Throughput optimality is established by the Lyapunov function $7$9 and a negative-drift argument: whenever the arrival-rate vector lies in the interior of the capacity region, the conditional expected drift is uniformly negative outside a bounded set, implying positive recurrence of the queueing Markov chain and stochastic boundedness of all queues (Tillman et al., 2022).

The paper also derives achievable request-rate regions for representative three-flow topologies. For three fully overlapping flows with $1.3$0 and $1.3$1, the constraints are

$1.3$2

$1.3$3

$1.3$4

For the orientation-aware case in which only head-to-tail aligned links are swapped, the bounds tighten to

$1.3$5

At $1.3$6, Monte-Carlo queue simulations over $1.3$7 slots with slope-threshold $1.3$8 validate the boundary (Tillman et al., 2022). In this setting, a switch is explicitly a queueing-and-routing fabric for entanglement requests rather than a simple physical gate.

4. High-voltage and cryogenic MEMS switch control

The MEMSDuino system addresses a different switching problem: room-temperature control of cryogenic MEMS RF switch networks. The system is a $1.3$9-inch rack-mount controller built around four modules: an Arduino microcontroller with button-reading ladder and NeoPixel LED driver interface, an Arduino shield board that translates GPIOs into relay-drive signals, a high-voltage generation stage, and an HV-switching stage. The relay board uses Comus $2.4$0-series SPST relays to route the $2.4$1 V line to the selected DSUB pin (Spietz et al., 6 Jan 2025).

The design is motivated by the observation that radio frequency cryogenic switches are a critical enabling technology for quantum information science, that solenoid-based switches have been used traditionally, and that a transition is being made to MEMS-based switches because of lower power dissipation, smaller size, and reduced risk from current pulses that can destroy expensive cryogenic amplifiers or cause electrostatic damage to devices. The paper states that MEMS switches require a $2.4$2-volt signal on the control lines, that built-in CMOS-based control electronics do not work at the cryogenic temperatures used in quantum information science, and that there is no currently available room-temperature control system with direct control of the switches (Spietz et al., 6 Jan 2025).

The controller allows manual switching via buttons on an LED-based indicator board and automatic switching via Python-based serial port commands to the Arduino. The firmware uses analog read of buttons via a resistor ladder on pin A0, a one-wire NeoPixel driver on digital pin D6, relay drive outputs on D2…D11, and serial at $2.4$3 baud listening for ASCII digits. The overview also records generic design relations such as

$2.4$4

and, for the Comus relays, $2.4$5, yielding $2.4$6 mA at $2.4$7 V. The exposition notes that the original paper does not report detailed timing or S-parameter data, but gives typical figures from Menlo’s MEMS devices and the relay-based driver: MEMS switching speed of $2.4$8s–$2.4$9 ms, Arduino command latency of less than $D_2$00 ms, power consumption of $D_2$01 mW per active coil, total idle power from the $D_2$02 V converter of less than $D_2$03 mW, RF isolation typically greater than $D_2$04 dB, insertion loss less than $D_2$05 dB at $D_2$06–$D_2$07 GHz, and reliability greater than $D_2$08 cycles at $D_2$09 K, with cryogenic lifetimes being qualified but showing no degradation after $D_2$10 cycles (Spietz et al., 6 Jan 2025).

A common narrow interpretation of switching hardware is that the switch itself is the only object of study. This system suggests a broader view: controller architecture, human interface, software API, high-voltage generation, and environmental placement at $D_2$11 K are also integral parts of the switching problem.

5. Skill switching in humanoid robotics

In humanoid robotics, “Switch” is the name of a hierarchical multi-skill system designed to enable seamless skill transitions at any moment. The framework has three components: a Skill Graph that augments demonstrated skill trajectories with kinematically similar cross-skill transitions and buffer states, a unified whole-body tracking policy trained by deep reinforcement learning on walks through this graph, and an online skill scheduler that performs graph search or nearest-neighbor planning for switching and recovery (Lau et al., 16 Apr 2026).

The Skill Graph is defined over a library $D_2$12 of demonstration sequences

$D_2$13

with graph

$D_2$14

Cross-skill edges are selected by nearest neighbor under the local-frame kinematic distance

$D_2$15

Deployment-time planning uses

$D_2$16

and when $D_2$17 is large, the system inserts $D_2$18 buffer nodes between $D_2$19 and $D_2$20 to preserve dynamic feasibility (Lau et al., 16 Apr 2026).

The whole-body tracking policy receives proprioceptive feedback and a graph-derived target,

$D_2$21

and is trained with PPO. The instantaneous reward includes an imitation term, a foot-ground contact reward,

$D_2$22

and action regularization. Training enhancements include a modified Reference State Initialization that samples start states at most $D_2$23 steps before a skill transition, buffer-aware imitation, and a curriculum that increases the fraction of augmented trajectories from $D_2$24 to $D_2$25 (Lau et al., 16 Apr 2026).

The online scheduler is triggered by initialization, user-commanded skill change, nearing the end of the current reference, or breach of a safety threshold. It performs a region-of-attraction check using $D_2$26, re-attaches directly if $D_2$27, engages emergency-stop if $D_2$28, and otherwise evaluates top-$D_2$29 entry candidates with

$D_2$30

where $D_2$31 is an approximate cost-to-go from reverse multi-source Dijkstra. The same graph can also support a nearest-neighbor recovery planner (Lau et al., 16 Apr 2026).

Experimental evaluation was conducted in MuJoCo and on a real Unitree G1 humanoid with $D_2$32 DoF and Jetson Orin NX. The motion library contained four skills plus acrobatic “Kung Fu” squats and back kicks. Difficulty levels were defined as Easy, Medium, and Hard with $D_2$33, $D_2$34, and $D_2$35 switches respectively, with $D_2$36 random trials per level under unperturbed and $D_2$37 N lateral push perturbations. The reported Skill Switching Success Rate for Switch (Base+SG+B+C) is $D_2$38 at all difficulty levels, whereas a single-skill Base policy stagnates at $D_2$39 and GMT drops from $D_2$40 as difficulty rises. Under perturbation, $D_2$41 m in Easy and $D_2$42 m in Hard (Lau et al., 16 Apr 2026). Here, switching denotes online graph-constrained replanning for whole-body behavior rather than mere concatenation of motion clips.

6. Tangible control interfaces and the SWITCH benchmark

The benchmark named SWITCH—“Semantic World Interface Tasks for Control and Handling”—extends the meaning of switch from device to evaluation substrate. It is an embodied, task-driven benchmark for tangible control interfaces such as light switches, appliance panels, and embedded GUIs, created to test perception, reasoning, action, and verification under egocentric RGB video input and device diversity (Lin et al., 20 Nov 2025).

Its first iteration, SWITCH-Basic, evaluates five complementary abilities: task-aware visual question answering, semantic UI comprehension, action generation, state-transition prediction, and result verification. Across these tasks, SWITCH-Basic comprises $D_2$43 multiple-choice questions drawn from $D_2$44 real-world egocentric video sequences covering $D_2$45 distinct devices. The paper further reports $D_2$46 fine-grained actions, $D_2$47 annotated states, and $D_2$48 unique $D_2$49 pairs. Automatic evaluation uses accuracy,

$D_2$50

with analogous forms for action-generation and verification accuracy (Lin et al., 20 Nov 2025).

The reported zero-shot findings show several recurring failure modes. Qwen3-VL-235B-Instruct reaches up to $D_2$51 on state-transition prediction in the Image-in-Question, Text-Choices setting, but drops to $D_2$52 when selecting among Image Choices, which the paper interprets as strong reliance on textual cues. Gemini 2.5 Flash shows Action Generation accuracy dropping from $D_2$53 in the IT setting to $D_2$54 in the VV setting, which the paper uses to illustrate under-utilization of visual and video evidence. Action Generation is identified as the hardest task, with models around $D_2$55–$D_2$56 accuracy, whereas UI Comprehension reaches $D_2$57–$D_2$58 and state-transition prediction exceeds $D_2$59. Verification planning scores above $D_2$60, but predicting the expected post-action state in Verification (b) drops to $D_2$61–$D_2$62 (Lin et al., 20 Nov 2025).

These results clarify a common misconception about switch handling in embodied systems. Interacting with a switch is not exhausted by identifying a button or issuing an actuation command. The benchmark explicitly separates semantic grounding, procedural action, causal prediction, delayed effects, and post-hoc verification. A plausible implication is that future work on switching in embodied intelligence will need tighter coupling among perception, planning, and verification than is common in static VQA or single-step action benchmarks (Lin et al., 20 Nov 2025).