RollArt System: Modular Advances

• RollArt System is a multi-domain framework combining distributed reinforcement learning, modular wearable dexterous manipulation, and rolling-phase OCT imaging.
• It leverages specialized hardware clusters, modular Roller Rings, and programmable phase ramp protocols to optimize scalability and performance.
• Empirical results show significant throughput gains in RL, high success rates in in-hand manipulation, and improved SNR in OCT imaging.

The term "RollArt System" refers to distinct, technically rigorous frameworks across three separate research domains: 1) disaggregated reinforcement learning infrastructure for LLMs, 2) modular wearable devices for dexterous in-hand manipulation, and 3) a rolling-phase regime for dynamic full-field OCT. Each instantiation of RollArt exemplifies domain-specific engineering and modeling, contributing foundational advances to its respective field.

1. Distributed RollArt System for Agentic RL Training

In high-throughput agentic RL, RollArt denotes a distributed runtime and resource-management paradigm explicitly tailored for large-scale, heterogeneous LLM post-training workloads. Unlike monolithic RL trainers, RollArt (Gao et al., 27 Dec 2025) partitions the reinforcement learning pipeline into specialized hardware pools that maximize throughput by matching workloads to hardware strengths.

Architectural Partitioning

The RollArt system comprises four principal resource clusters:

• Training cluster: Compute-optimized GPUs (e.g., NVIDIA H800) execute distributed gradient updates (Megatron).
• Inference cluster: Bandwidth-optimized GPUs (e.g., NVIDIA H20) serve LLM generation via proxy layers (vLLM/SGLang-based).
• CPU cluster: Kubernetes-managed stateful simulators (EnvManager), with persistent environment state per instance.
• Serverless pool: Stateless reward evaluation mapped to Function Compute.

Data flows asynchronously: an EnvManager dispatches generate requests to LLMProxy (prefill/decoding), applies actions to the environment, and forwards completed trajectories for reward evaluation. Trajectories are batched in a SampleBuffer (Ray ObjRefs), from which policy training clusters pull asynchronously, with model weights synchronized via Mooncake.

Core Design Principles

Three design commitments underlie RollArt’s efficacy:

1. Hardware-Affinity Workload Mapping: Annotation-based scheduling routes compute-intensive and decoding-bound tasks to optimal GPU types via explicit affinity mappings. Prefill-heavy tasks are empirically faster on H800; decoding-heavy workloads favor H20.
2. Fine-Grained Asynchrony: RollArt eschews batch synchronization by running trajectory-level event loops in the rollback and training stages. Suspension-resume semantics minimize resource underutilization and enable managed version staleness (parameterized by $\alpha$).
3. Statefulness-Aware Computation: Stateful components (environment, training) are containerized with affinity; reward computation, being stateless, is elasticized via serverless workers. This division maximally utilizes available hardware and reduces rollout latency.

Algorithms and Performance

Key runtime features include a scheduler filtering workers by affinity, trajectory-level event processing per environment, and asynchronous model weight propagation. The end-to-end step time $$T_{\text{step}} \approx \max(T_{\text{rollout}}, T_{\text{train}}+T_{\text{sync}})$$. Asynchronous rollout allows RollArt to hide training and synchronization in rollout latency.

Reported results include step time reduction up to 2.05$\times$, throughput efficiency gains of 2.65–4.58$\times$, and robust scaling to >3,000 GPUs, with environment reset success >99.99% over a week-long training run (Gao et al., 27 Dec 2025).

2. RollArt (Roller Ring) Modular Manipulation System

RollArt, as a hardware and control solution for in-hand manipulation, consists of modular, wearable Roller Rings (RRs) that actively induce object reorientation and translation without requiring finger repositioning (Webb et al., 2024). The theoretical model unifies non-holonomic contact kinematics with closed-loop actuation.

Mechanical and Actuation Design

Each RR module integrates:

• Physical Geometry:

Outer diameter ≈ 30 mm (adapted for finger or robot digit sizes), active belt width ≈ 18 mm.

• Actuation:

Micro N20 DC gearbox (210:1) drives a timing-belt loop; closed-loop with 12 CPR encoder.

• Modularity:

Universal 3D-printed guide tracks and conformable affixing sleeves (CASM) enable fit for diverse appendage geometries (human, Yale Model O).

• Control Surface Configuration:

Each ring’s belt axis is tilted by an adjustable angle (typically 30°), ensuring non-parallel actuation for dexterous, multi-axis manipulation.

Schematic (as per source):

\begin{figure}[h] \centering \begin{tikzpicture}[scale=1] \draw[thick] (0,0) -- (0,2) node[above]{Finger\,1}; \draw[thick] (2,0) -- (2,2) node[above]{Finger\,2}; \draw[blue] (0,1) circle (0.6) node[right=5pt]{RR%%%%4%%%%}; \draw[red] (2,1) circle (0.6) node[right=5pt]{RR%%%%5%%%%}; \draw[->, thick, blue] (0,1) -- ({0.6*cos(30)},{1+0.6*sin(30)}) node[above right]{%%%%6%%%%}; \draw[->, thick, red] (2,1) -- ({2+0.6*cos(120)},{1+0.6*sin(120)}) node[above left]{%%%%7%%%%}; \draw[<->] (0.2,1.2) arc (60:90:0.5) node[midway,right]{%%%%8%%%%}; \end{tikzpicture} \caption{Schematic of two RRs on parallel fingers with belt axes tilted by %%%%9%%%%.} \end{figure}

Kinematic and Differential Motion Model

The RollArt RR system implements a non-holonomic control scheme, employing differential kinematic constraints:

• For $N$ contacts (rings), the net angular velocity is:

$$\omega_* = \frac{\sum_{i=1}^N k_i\, \omega_i}{\sum_{i=1}^N k_i}$$

where $k_i$ scales tangential force from belt slip. For general shapes, torque-weighted by contact distance $r_i$:

$$\omega_* = \frac{\sum_{i=1}^N k_i\, r_i\, \omega_i}{\sum_{i=1}^N k_i\, r_i}$$

• Two RRs with non-coincident axes provide a complete, non-holonomic basis for SO(3) in-hand reorientation. With additional RRs, the control cone expansion improves local manipulability and reduces detour travel.
• Object twist constraints (Pfaffians) ensure that finger/object relative slips are actively controlled instead of relying on passive rolling.

Dexterity and Empirical Performance

Extensive validation on both robotic (Yale Model O) and human hands demonstrates:

Object	Operation	Success [%]	Max Speed [m/s]
A–F Sphere	Reorient	98	0.025
Cube	Reorient	94	0.018
Pringles Tube	Translate	90	0.015
Cheese Toy	Rotate+Translate	92	0.020

Tasks include arbitrary-object pose reorientation and translation, with empirical rotation speeds up to 0.02 m/s and success rates consistently above 90%. The modular architecture allows robust trajectory correction through adaptive control of belt speeds.

Applications and Extensions

Current and plausible future applications include field robotics, rehabilitation assistive exoskeletons, AR/VR haptics, and industrial retrofits for dexterous manipulator end-effectors. Potential improvements include compliance engineering (flexures), additional degrees of belt freedom, surface friction regulation, and incorporation of MPC or learning-based force/encoder adaptation (Webb et al., 2024).

3. Rolling-Phase (RollArt) DFFOCT Regime

Within full-field optical coherence tomography (FFOCT), RollArt (termed Rolling-Phase or RP-DFFOCT) denotes a novel acquisition and demodulation protocol employing a slow, programmable reference-arm phase ramp to provide simultaneous, artifact-free extraction of static and dynamic contrast (Monfort et al., 14 Jan 2025).

Optical and Hardware Configuration

The RP-DFFOCT implementation utilizes:

• Low-coherence LED: λ₀ = 810 nm, Δλ = 25 nm (Thorlabs M810L3).
• Linnik interferometer: Balanced sample/reference objectives (Olympus UPLSAPO30XSIR, 30×, 1.05 NA).
• Reference mirror actuation: PZT (Thorlabs PK44M3B8P2), driven by DAC (NI-9263), applies a linear phase ramp $\phi_{\text{ref}}(t)$ over M·2π radians.
• CMOS camera (Adimec Q-2HFW): Frame-synchronized to PZT via FPGA/DAC, resolution typically 512×512 ROI.

Mathematical Model and Signal Extraction

The formalism considers:

• Interferometric intensity (neglecting envelope):

$$I(t) = I_r + I_s + I_{\text{inc}} + 2\sqrt{I_r I_s} \cos[\phi_0 + \phi_s(t) + \phi_{\text{ref}}(t)]$$

• The imposed phase ramp:

$$\phi_{\text{ref}}(k\Delta t) = 2\pi M k / N,\ \ k = 0,1,...,N\!-\!1$$

with $N = 512$, $M = 1$ or $2$ (total phase sweep of 2π or 4π), $f_{\text{frame}} = 100$ Hz.

• Static contrast: Extracted as the Fourier coefficient at the carrier frequency $f_0 = M f_{\text{frame}} / N$:

$$I_{\text{static}}(x,y) = |\text{FFT}_t \{ I(x,y,t) \}(f_0)|$$

• Dynamic contrast metric: Time-averaged absolute difference,

$$B(x,y) = \langle | I(t+\Delta t) - I(t) | \rangle_t \simeq \frac{8}{\pi}\sqrt{I_r I_s} \langle | \Delta\phi_s | \rangle_t$$

• Combined HSB maps visualize dynamic (brightness), frequency content (hue), and power spectral density (saturation) in a single acquisition.

Performance and Artifact Suppression

RollArt (RP-DFFOCT) achieves:

• $\sim4\times$ improvement in SNR for dynamic contrast relative to standard DFFOCT (same frame count).
• Suppression of spurious fringe artifacts around high reflectors by >90%, and global speckle-contrast reduction of $\sim$20%.
• Axial resolution: $\sim1\,\mu$m (coherence limit), transverse: $\sim0.4\,\mu$m.
• Reduced acquisition requirements ($\sim$¼ native frames) for equivalent discriminative mapping.
• Artifact-free single-shot extraction of both static and dynamic maps.

Experimental imaging of macaque retinal explant demonstrates elimination of ring artifacts in photoreceptors and enhanced visualization of individual nuclear and axonal activity—previously obscured by speckle/fringe artifacts (Monfort et al., 14 Jan 2025).

4. Mathematical and Dynamical Foundations of Rolling Nonholonomic Systems

Rigorous mathematical models for rolling rigid bodies underpin both the hardware (Roller Ring) and control-theoretic (rolling-phase) implementations. The semi-symplectic formalism developed by Patrick (Patrick, 2017) formalizes configuration manifolds, shape operator constraints, and first-order ODEs subject to non-holonomic, no-slip rolling:

• Configuration: $$q = (A,s,x) \in SO(3) \times \mathcal{S} \times \Sigma$$, holonomic constraint $A n_{\mathcal{S}}(s) = n_{\Sigma}(x)$.
• Non-holonomic rolling constraint: $\dot x = A \dot s$, $$\Lambda(s,x)\dot s = \Omega \times n_{\mathcal{S}}(s)$$, with $\Lambda$ the differential shape-operator mismatch.
• Semi-symplectic ODEs preserve a modified energy, enabling reliable open-loop trajectory generation and geometric feedback synthesis for rolling actuation.

Integration of these models into modular hardware actuation (Roller Ring system) and control (RollArt DFFOCT) enables energy-conserving, constraint-respecting manipulation and imaging.

5. Impact, Applications, and Prospective Developments

RollArt frameworks across domains share a unifying emphasis on modularity, decoupling of heterogeneous subsystems, and the exploitation of non-holonomic constraints to expand the operational envelope of their respective technologies.

• In agentic RL: RollArt enables the scaling of LLM-driven agents to unprecedented model and cluster sizes. The disaggregation paradigm, hardware-affinity mapping, and asynchrony are likely to influence next-generation RL engineering.
• In robotic manipulation: The hardware RollArt system offers a direct path to retrofit manipulators for high-dexterity, in-hand tasks without demanding redesign or high DOF actuation. Extensions to compliance and advanced control are under investigation.
• In optical imaging: The rolling-phase protocol generalizes to mitigate phase noise in other interferometric and Fourier-domain imaging techniques, enhancing sensitivity and spatial discrimination.

This suggests that RollArt conceptually guides future efforts to partition complex, interdependent systems—whether computational, mechanical, or optical—into modular, asynchronously coordinated subunits to maximize performance, flexibility, and robustness.