CENIC: Convex Error-controlled Numerical Integration for Contact

AI summary

Overview

Research area: Robotics simulation, specifically contact-rich multibody dynamics (contact modeling, dexterous manipulation, simulation and animation).

Technical level: Advanced. The paper assumes familiarity with multibody dynamics, differential-algebraic equations, error-controlled integration, and convex optimization-based time-stepping.

Scope: CENIC is a new continuous-time, error-controlled numerical integrator for contact-rich robotic simulation that combines convex time-stepping with adaptive step-size control to achieve accuracy guarantees at real-time speeds.

What This Paper Is About

Most robotics simulators advance time in fixed discrete steps, which forces users to choose a time step that is either too large (producing non-physical artifacts like interpenetration, tunneling, and impulse jitter) or too small (forcing slow simulation). The authors introduce CENIC, a continuous-time integrator that automatically adjusts the time step to hit a user-specified accuracy, using a convex time-stepping formulation (irrotational contact fields, or ICF) so that the underlying optimization converges for any step size. The goal is to blend the speed and scalability of discrete-time simulators with the mathematical rigor and accuracy guarantees of engineering-grade error-controlled integration.

Key Contributions

1. First error-controlled integrator for contact-rich multibody dynamics tailored to robotics. CENIC builds on convex-optimization-based time-stepping and irrotational contact field (ICF) theory to create a convex error-controlled integration scheme.

2. Accuracy, consistency, and convergence guarantees. Users specify a desired accuracy rather than a fixed time step; as accuracy is tightened, the numerical solution converges to the true continuous-time trajectory, and the convexity of the underlying problem yields convergence guarantees independent of the current time step.

3. Engineering-grade contact and friction models. CENIC is built on continuous-time compliant contact models (point contact and hydroelastic contact, with Hunt & Crossley dissipation) and can rigorously model both static and dynamic friction regimes, unlike the discrete-time schemes typically used in robotics.

4. Practical performance optimizations and modularity. Arbitrary controllers and external systems are embedded in the convex formulation, and the authors demonstrate techniques such as selective Hessian reuse and adaptive convergence tolerances. The paper reports that CENIC outperforms traditional error-controlled integrators by orders of magnitude, and that overall simulation times can be faster than discrete-time methods.

Main Findings

• Real-time performance with guarantees: CENIC runs at fast real-time rates comparable to discrete-time robotics simulators like MuJoCo, Drake, and Isaac Sim, while also providing guarantees on accuracy and convergence.

• Eliminates contact artifacts in a challenging scenario: In the bimanual manipulation scenario of Fig. 1 (a teleoperated robot dropping a spatula, spoon, and mug into a thin-walled bin, then dumping contents onto a thin-wired dish rack), discrete time-stepping methods struggle with passthrough, rattling, and unstable contact forces. CENIC eliminates these artifacts while completing the simulation at a real-time rate over 300%.

• Addresses three failure modes of conventional integrators: (1) Newton–Raphson iterations in implicit schemes often diverge, forcing step rejection; CENIC's convexity guarantees convergence for any step size. (2) Friction instabilities force conventional integrators to monitor velocities and take small steps to resolve impact transitions; ICF yields well-converged tangential velocities and coherent friction forces at any time step. (3) Standard implicit integrators require costly geometry queries at each Newton iteration; CENIC needs only two geometry evaluations per step.

• Tight stiction regularization: CENIC works with a stiction tolerance of v_s ≈ 0.1 mm/s and a friction transition width of Δ = 10. For reference, the paper notes Simbody uses a default stiction tolerance of 1 cm/s, while Hyfydy and MSC Adams use 10 cm/s.

• Consistency of the underlying scheme: The symplectic IMEX scheme used by ICF is first order, with first-order approximations of compliance and friction, and is consistent with the continuous model — the discrete scheme recovers exact trajectories of the DAE in the limit of vanishing step size.

• Error control details: The paper describes a first-order step-doubling method (Section V-A) and a second-order scheme (Section V-B). The authors note that for a typical near-rigid estimation parameter, β = 0.1 is typical and delivers tight constraint enforcement in practice. They also reference the Dormand-Prince integrator, which computes one estimate with order 5 and another with order 4, giving a fifth-order error estimate (p = 5).

• Not reported in the available content: The detailed experimental results (Section VII) — including specific comparisons against MuJoCo, Drake, and Isaac Sim — are not included in the truncated paper content provided. The paper's full numerical benchmarks, timing tables, and convergence plots therefore cannot be summarized here.

Methodology in Plain English

The authors model contact in continuous time, using compliant contact models (point contact and hydroelastic contact, both with Hunt & Crossley dissipation) so that normal forces are smooth functions of state. Friction is regularized with a smooth function that transitions between static and dynamic friction regimes, controlled by a single stiction tolerance velocity parameter.

Rather than solving the resulting stiff differential-algebraic system with a conventional stiff integrator (which stalls on near-rigid contact and tight friction regularization), CENIC borrows a convex time-stepping formulation from ICF theory. In this formulation, velocities at the next step are the minimizer of a convex cost function, so the implicit step always has a unique solution and converges regardless of step size. CENIC wraps this convex step inside a standard error-controlled integration loop: it attempts a step, computes two estimates of the next state, uses their difference to estimate local truncation error, and adjusts the step size to meet a user-specified accuracy — rejecting and retrying steps when the error is too large.

External components such as controllers, planners, and learned policies are abstracted as a separate dynamical system in feedback with the multibody system and integrated implicitly as well, which the authors note is effective even for stiff feedback dynamics. Limit and holonomic constraints are handled by regularizing them in the same convex formulation, with stiffness and damping set according to a "near-rigid" estimation so that constraints tighten as the time step shrinks. The paper also describes optimizations such as selective Hessian reuse and adaptive convergence tolerances to improve speed.

Why This Matters

Impact on research: The paper argues that discrete-time simulators — including Bullet, MuJoCo, Isaac Sim, RaiSim, Genesis, and Drake — trade accuracy for speed, and that the gliding/hydroplaning artifacts of Anitescu-based convex relaxations (which, for compliant formulations like MuJoCo and SAP, do not vanish as the time step is reduced) force users to choose models further from reality. CENIC aims to provide a foundation for scalable, physically consistent simulation with provable accuracy and convergence guarantees, potentially enabling tighter sim-to-real alignment.

Real-world applications:

• Dexterous manipulation, including the bimanual scenario in Fig. 1 with heavy arms, thin objects, and stiff joint controllers.
• Locomotion and reinforcement learning, which require simulation speeds at or above real time.
• Interactive teleoperation and model predictive control, which need fast simulation throughput.
• Large-scale contact-rich environments such as full factory floors or domestic environments, where the number of contacts and degrees of freedom grows.

Industry relevance: Commercial multibody dynamics tools in aerospace, automotive, and mechanical engineering (MSC Adams, Siemens Simcenter Motion, RecurDyn) prioritize accuracy and support V&V/UQ frameworks, FEA integration, and CAD, but the paper states they are largely unsuitable for robotics because high stiffnesses and tight friction regularization produce prohibitively stiff dynamics. CENIC aims to bring that engineering-grade reliability into robotics-grade simulation speed.

Future Directions

• Detailed empirical evaluation: the full experimental results comparing CENIC against discrete-time simulators and conventional stiff integrators are not included in the provided content, leaving open the question of how the reported real-time performance holds across the full benchmark suite.
• Extending the approach beyond the two schemes described (first-order step-doubling and second-order) to higher-order error estimates and broader integrator families.
• A more thorough treatment of non-holonomic constraints, which the paper sets aside because their main use case (rolling) is already captured by the contact model.
• Scaling and verification/validation (V&V) and uncertainty quantification (UQ): applying the aerospace/automotive-style formal frameworks to robotic contact simulation, which the paper identifies as a gap in current robotics tools.

Target Audience

Robotics researchers and engineers working on simulation, contact modeling, manipulation, and locomotion; developers of physics engines and multibody dynamics software; and reinforcement learning or control practitioners who need fast, trustworthy contact simulation. Readers without a background in numerical integration or convex optimization will find the mathematical sections (Section IV and V) demanding.

Authors’ abstract

State-of-the-art robotics simulators operate in discrete time. This requires users to choose a time step, which is both critical and challenging: large steps can produce non-physical artifacts, while small steps force the simulation to run slowly. Continuous-time error-controlled integration avoids such issues by automatically adjusting the time step to achieve a desired accuracy. But existing error-controlled integrators struggle with the stiff dynamics of contact, and cannot meet the speed and scalability requirements of modern robotics workflows. We introduce CENIC, a new continuous-time integrator that brings together recent advances in convex time-stepping and error-controlled integration, inheriting benefits from both continuous integration and discrete time-stepping. CENIC runs at fast real-time rates comparable to discrete-time robotics simulators like MuJoCo, Drake and Isaac Sim, while also providing guarantees on accuracy and convergence.