RollArt (RollArc): Unified Rolling Dynamics

• RollArt (RollArc) is a multidisciplinary framework that combines mathematical modeling, physical rolling dynamics, and generative art techniques for precise control and kinetic aesthetics.
• It employs inverse-rolling methods, nonholonomic constraints, and screw theory to solve complex trajectory planning problems and optimize edge-rolling manipulation.
• Applications include kinetic art installations, physics demonstrations, and high-performance distributed reinforcement learning systems that enhance robotic and computational efficiency.

RollArt (RollArc) refers to a collection of mathematical, physical, algorithmic, and infrastructural frameworks centered on objects rolling along prescribed curves, generative art techniques based on rolling kinematics, manipulation planning, and large-scale reinforcement learning pipelines explicitly exploiting rolling/trajectory asynchrony. Across these domains, RollArt/RollArc structures are unified by the rigorous modeling and technological deployment of rolling, slipping, and path-following dynamics, realized algorithmically and physically. The following sections detail key theoretical pillars, characteristic methodologies, and principal domains of application.

1. Rolling Kinematics and Inverse Path Construction

The mathematical essence of RollArc lies in the inverse-rolling problem: given a desired planar trajectory $\gamma: [s_0, s_1] \to \mathbb{R}^2$, construct a strictly convex, $C^2$-smooth body $\mathcal{B}$ whose boundary $\partial\mathcal{B}$, upon rolling without slip, causes the body’s center to trace $\gamma$ and—after a minimal integer $n$ of curve repeats—regains its original orientation. The central mechanism is the local no-slip constraint: at each moment the instantaneous center of rotation of $\mathcal{B}$ coincides with the point of contact, yielding $ds = R_\mathrm{eff}(\theta) d\phi$, where $R_\mathrm{eff}$ is the local effective radius at angle $\theta$.

Solving the inverse problem proceeds via the support function $C^2$0 of $C^2$1, obeying the ODE:

$C^2$2

where $C^2$3 is the geodesic curvature of $C^2$4 and $C^2$5 is defined by $C^2$6 (Eckmann et al., 2024). The solution determines a unique convex roller, up to translations and rotations, that robustly traverses $C^2$7. The minimal closure period $C^2$8 follows from the total tangent rotation $C^2$9: if $\mathcal{B}$0 (lowest terms), then closure occurs after $\mathcal{B}$1 repeats.

Key algorithmic steps:

1. Compute the curve’s curvature and cumulative tangent rotation.
2. Numerically invert the $\mathcal{B}$2 map.
3. Solve the banded linear (or spectral) system for $\mathcal{B}$3 with periodicity and zero-drift constraints.
4. Reconstruct boundary points via the standard support-function formula: $\mathcal{B}$4.
5. Export the boundary for manufacturing or simulation.

For classical shapes—circles (pure cycloids), ellipses (egg-shaped rollers), etc.—explicit construction verifies agreement with the general theory.

2. Rolling Dynamics, Nonholonomic Constraints, and Geometric Mechanics

The differential-geometric framework for rolling motion is centered on nonholonomic constraints and semi-symplectic reduction. The configuration space for a rigid body $\mathcal{B}$5 rolling on surface $\mathcal{B}$6 is $\mathcal{B}$7, with contact and normal-incidence constraints: $\mathcal{B}$8 The rolling (no-slip) constraint is $\mathcal{B}$9, where $\partial\mathcal{B}$0 is the body-frame velocity and $\partial\mathcal{B}$1 is angular velocity. The intrinsic rolling constraint becomes

$\partial\mathcal{B}$2

with $\partial\mathcal{B}$3 defined in terms of the Weingarten maps (shape operators) $\partial\mathcal{B}$4 and $\partial\mathcal{B}$5 (Patrick, 2017).

Dynamics are governed by the left-invariant Lagrangian plus gravity, with the reduced inertia $\partial\mathcal{B}$6, leading to the semi-symplectic two-form $\partial\mathcal{B}$7 and associated energy $\partial\mathcal{B}$8. The equations of motion on $\partial\mathcal{B}$9 (the no-slip constraint distribution) are: $\gamma$0 and

$\gamma$1

These formulae enable the simulation and analysis of arbitrary rigid bodies rolling along arbitrary smooth tracks.

3. Algorithmic Edge-Rolling and Manipulation Planning

RollArc (in manipulation) designates a planning and control methodology for prehensile edge-rolling of curved-edge objects (e.g., cylinders) along arbitrary paths using sequences of constant screw motions and optimization. Each rolling segment realizes a pure rotation about the instantaneous axis at the current contact, modeled using screw theory and dual quaternions: $\gamma$2 where $\gamma$3 codes rotation and $\gamma$4 codes translation but edge-rolling uses $\gamma$5 (pure rotation). The continuous path is discretized into $\gamma$6 segments of length $\gamma$7, each realized by rotations about endpoints, $\gamma$8. To transition between non-colinear segments, ScLERP pivots are executed (Boroji et al., 2024).

Optimization, subject to robot and joint constraints, minimizes deviation from a straight trajectory under bounds on pivot angle and roll length. Algorithms solve for segment lengths and angles, sequentially executing roll and pivot primitives, then synthesizing the full joint trajectory by inverse kinematics.

Physical validation (e.g., with a Franka Panda arm rolling a rigid cylinder) demonstrates that RollArc provides higher maneuverability and more flexible path following than sliding-only or pure-pivot strategies, with negligible slip at sufficient discretization.

4. Geometric and Artistic Realizations: Spirographics, Platonicons, and Caustics

The RollArt system generalizes rolling to create and explore parametric plane curves (trochoids, epitrochoids, cycloids, ellipses) by combining controlled rolling and sliding of circles along circles or lines, physically instantiated via devices such as the Mechanical Oscilloscope (MO) and algorithmically by Virtual Rotating Circles Technique (VRCT) and Virtual Sliding Simulator (VSS) (Arbab et al., 2024). The core parametric forms unify all such curves: $\gamma$9 for rolling/sliding on the outside, with modifications for inside rolling or line-based cycloids.

By tuning the spin rates $n$0, radii, and slide parameters $n$1, the system can generate the entire space of classical spirographic curves, continuously interpolating between cycloids, trochoids, and ellipses. This provides a generative paradigm for kinetic and visual art installations, supporting both algorithmic and mechanical implementations.

Platonicon-based RollArt extends these ideas into three dimensions, generating developable kinetic sculptures by "dressing" Platonic solids with modules based on conical segments. Each module corresponds to a cone of semi-apex equal to the dihedral angle of the dual solid, ensuring global developability. The rolling trajectory of the assembled Platonicon is a closed meandering path with fixed height of the center of mass and total net displacement zero after a complete revolution. Artistic guidance includes color segmentation, module edge-highlighting, and modular fabrication (Seaton et al., 2020).

In optical caustics, the caustic envelope of reflected rays from a curve $n$2 and radiant $n$3 is precisely parametrized by the roulette of a family of rolling circles $n$4 along their second envelope $n$5, with the marked point $n$6 traced as

$n$7

where $n$8 encodes the geometric reflections and $n$9 is the focal radius (Boyle, 2014).

5. Physics Demonstration and Educational Implementations

RollArt/RollArc is exemplary in physics education, demonstrating conservation of energy and dynamics of rolling bodies through experiments such as a paint-coated small sphere rolling down an exercise ball (Phan-Budd, 2018). Key physics:

• Energy conservation yields the velocity $\mathcal{B}$0 for the rolling ball.
• The release angle $\mathcal{B}$1 at which contact is lost arises from balancing centripetal and gravitational forces.
• The observed paint tracks confirm the expected arc lengths and provide artistic variants by varying release angle, surface tilt, and paint thickness.
• The geometry of the paint tracks, their arclengths, and curvature (projections as circles or ellipses) are explicitly quantified, enabling both physical understanding and artistic composition.
• Guidelines are given for parameter selection, color arrangements, and setup, fostering accessibility for both scientific demonstration and art.

6. Large-Scale Distributed Reinforcement Learning: The RollArc System

In the context of agentic reinforcement learning (RL) for LLMs, RollArc denotes a high-performance disaggregated infrastructure for scaling out RL workloads across heterogenous compute resources (Gao et al., 27 Dec 2025). The design motivation stems from the heterogeneous nature of agentic RL training:

• The rollout phase interleaves compute-bound prefill (high TFLOPS) with bandwidth-bound decoding and thousands of stateful, CPU-heavy environment simulations.
• The reward stage is dominated by stateless, low-utilization evaluations, unsuitable for exclusive GPU scheduling.
• The training stage requires tight GPU connectivity for high-performance gradient steps.

RollArc achieves throughput maximization and resource optimization using three core principles:

P1. Hardware-Affinity Mapping: Sub-tasks are tagged for preferred hardware (e.g., H800 vs. H20 GPUs), with the scheduler minimizing trajectory latency via $\mathcal{B}$2, achieving up to $\mathcal{B}$3 step speedup over homogeneous allocations.

P2. Fine-Grained Asynchrony: RollArc orchestrates execution at the trajectory level, eliminating global batching barriers and resource idling. Asynchronous rollout and training, combined with explicit trajectory version staleness bounds ($\mathcal{B}$4), cut resource bubbles by up to 90%, yielding up to $\mathcal{B}$5 speedup versus batch RL.

P3. Statefulness-Aware Computation: Stateless reward computation is offloaded via serverless infrastructure, auto-scaling to match evaluation demand and increasing GPU reward utilization from $\mathcal{B}$6 to $\mathcal{B}$7.

Empirical deployment on over 3,000 GPUs (Alibaba) demonstrates $\mathcal{B}$8–$\mathcal{B}$9 reduction in time-to-score and $ds = R_\mathrm{eff}(\theta) d\phi$0 throughput over synchronous baselines when training massive MoE LLMs. Bottlenecks, limitations, and open questions include dynamic hardware mapping, policy staleness-versus-throughput in asynchrony, and adaptive multi-tenant communication (Gao et al., 27 Dec 2025).

7. Synthesis: RollArt/RollArc as a Unified Framework

Across all domains, RollArt/RollArc structures exploit the geometry and dynamics of rolling—pure or in combination with sliding, pivoting, or asynchrony—to achieve precise trajectory control, generative art, efficient computation, and educational clarity. Methodological commonalities include:

• Unification of kinematics via support functions, screw theory, or hardware mapping abstractions.
• Algorithmic realization as explicit ODE/PDE solvers, sequential rolling-pivoting planners, or asynchronous distributed runtimes.
• Physical realization in demonstration and kinetic sculpture, exploiting geometric developability and controlled fabrication.

RollArt thus serves as a paradigmatic bridge linking geometric control, computational geometry, artistic composition, and large-scale learning system design.