cs.RONov 11, 2025

CENIC: Convex Error-controlled Numerical Integration for Contact

Authors: Vince Kurtz, Alejandro Castro

Read PDF
Ask AI
Cite

Copy citation

DOINot available for this paper

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 variable-step 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.

Figures & tables

Fig. 1 : A challenging bimanual manipulation scenario in simulation (top) and on hardware (bottom). A teleoperated robot drops several objects (spatula, spoon, mug) into a thin-walled bin. It then dumps the contents of the bin on a dish rack composed of thin wires. This scenario is particularly challenging: complex geometry, thin objects, heavy arms, and stiff joint controllers create a wide range of time scales that pose significant numerical difficulties. 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%.
Fig. 2 : CENIC’s regularized friction model captures both static ( μs\mu_{s} ) and dynamic ( μd\mu_{d} ) friction regimes. Our default stiction tolerance is vs=0.1 mm/sv_{s}=0.1\,\text{mm/s} , with a transition width of Δ=10\Delta=10 .
Fig. 3 : In practice, robotics simulations involve multibody dynamics connected to controllers, planners, and other components. We abstract these elements as a single external dynamical system. Unlike discrete formulations, CENIC integrates external systems implicitly, an effective strategy even for stiff feedback dynamics.
Fig. 4 : A simple nonlinear controller ( 30 ) induces a series of stable and unstable equilibria in a torque-controlled pendulum. CENIC’s linearization strategy integrates the controller implicitly in the stable (shaded, negative slope, negative feedback) regions. To maintain convexity, we fall back to an explicit scheme in the unstable regions. Enforcing convexity everywhere while performing implicit integration whenever possible enables faster and more reliable simulation of complex nonlinear controllers, see Sec. VII-D1 .
Fig. 5 : Vertical position and velocity of a ball dropped on a table, showing differences between solutions computed with a 1 ms (solid line) and a 5 ms (dashed line) time step. Velocity error spikes, but position error does not.
 Discrete Solvers [ 1 , 2 , 3 , 4 , 5 ]Error-Controlled Implicit Integrators [ 15 ]CENIC
Simulate Stiff Frictional ContactFastVery SlowFast
Guaranteed AccuracyNoYesYes
ConvexSometimesNoYes
Geometry QueriesOne per stepOne per iteration (many per step)Two per step
Reduce TimestepNeverFor accuracy and convergence failureFor accuracy only
Implicit External SystemsNoYesYes
TABLE I : Comparison of key characteristics of different simulation strategies. CENIC brings together the best of classical error-controlled integration and discrete solvers often used for robotics simulation.
Fig. 6 : Solver convergence during hard clutter simulation, illustrating the impact of Hessian reuse. Red dashed lines indicate a recomputed Hessian, while light grey lines separate time steps. ( 36 ) allows us to reuse the same Hessian when possible (first ten steps). When necessary, a few fresh factorizations enable rapid Newton-type convergence (final step).
NameDegrees of FreedomMax. Num. Contacts
cylinder61
gripper73
franka162
soft clutter12031
hard clutter12077
TABLE II : Scale summary (degrees of freedom and maximum number of simultaneous contacts) for the test cases shown in Fig. 7 .
Fig. 7 : Simulation test cases, ranging from least to most complex. These test cases highlight different challenges in robotics simulation: sharp corners ( 7(a) ), stable force closure ( 7(b) ), stiff nonlinear controllers ( 7(c) ), many degrees-of-freedom ( 7(d) ), and complex contact interactions ( 7(e) ).
Fig. 8 : Contact force magnitude for the static gripper example at accuracy εacc=10−3\varepsilon_{\text{acc}}=10^{-3} . CENIC (top) produces smooth, stable contact forces, while RK3 (bottom) produces noisy, unstable results. Deceivingly, positions might look visually accurate with RK3, until later in the simulation when RK3 mistakenly drops the object.
Fig. 9 : Energy conservation error (percent of energy lost after 10 seconds) for a 0.1 kg bouncing ball with zero dissipation. The ball is relatively soft (contact stiffness 10310^{3} N/m) and bounces 11 times in the 10 second simulation. Potential energy is defined such that total energy is zero when the ball is at rest on the ground. The trapezoid method (Sec. V-B ) shows second-order convergence, while step-doubling (Sec. V-A ) is first order.
Fig. 10 : Work-precision plots for hard clutter (lower is better). Despite requiring only two convex solves per step versus three for step-doubling, the trapezoid method’s lack of stability hinders performance.
Fig. 11 : Work-precision plots for the test cases in Fig. 7 , using error control on the full state x\mathbf{x} (top) and on positions q\mathbf{q} only (bottom). Lower is better. Wall times are normalized to one second of simulation time. Missing data points indicate solver failure or timeout after 100 seconds (real-time rate <1%<1\% ). CENIC performs particularly well with many contacts and at loose accuracies, where convergence failures force other implicit schemes to take small steps.
Fig. 12 : Simulator speed comparison: wall times to complete a 10-second (top) and 3-second (bottom) simulation. Three mesh bunnies (top) and an unactuated quadruped robot (bottom) are dropped on the ground. CENIC is competitive with state-of-the-art robotics simulators in both fixed-step and error-controlled mode. Note that these benchmarks measure speed only, not simulation quality. Also note that the GPU simulators are optimized for simulating many parallel scenes, but these tests use only a single scene.
Fig. 13 : Number of time steps required to simulate the franka task for 10 seconds with different controller stiffnesses ( =Kp=Kd=K_{p}=K_{d} ) and accuracy εacc=0.01\varepsilon_{\text{acc}}=0.01 . With a standard explicit treatment, step size must decrease to keep stability under control. In contrast, our proposed implicit scheme (Sec. V-C ) is stable even with extremely stiff controllers.
Fig. 14 : Suction gripper after 0.5s of simulation. For discrete time-stepping, we use δt=0.01\delta t=0.01 s. For CENIC, we request accuracy εacc=10−3\varepsilon_{\text{acc}}=10^{-3} . Large forces pulling the box into the gripper lead to massive passthrough artifacts in the discrete-time simulation. CENIC resolves the dynamics correctly with a combination of error control (Section V-A ) and implicit treatment of the gripper force model (Section V-C ).
Fig. 15 : Tangential velocities and friction forces for a 1 kg box on an oscillating conveyor belt, demonstrating transitions between static ( μs=1.0\mu_{s}=1.0 ) and dynamic ( μd=0.5\mu_{d}=0.5 ) friction. An accuracy of 10−510^{-5} (left) shows the box “catching” in stiction at low velocities. A tight accuracy ( 10−810^{-8} , right) is needed to capture the impulsive transition from sliding to stiction. Existing convex time-stepping methods are unable to capture any of these effects, as they only support a single coefficient of friction.
Fig. 16 : Undesirable artifacts from simulating the dish rack example (Fig. 1 ) with a too-large time step (fixed-step mode) or a too-loose accuracy (error control mode). Tightening the accuracy is a more efficient means of eliminating these artifacts than reducing the time step, as shown in Table III
Accuracy10−110^{-1}10−210^{-2}10−3\mathbf{10^{-3}}10−4\mathbf{10^{-4}}
Real-time rate218%463%342%132%
ArtifactsYes ( 16(d) )Yes ( 16(e) )NoNo
Time Step (s)0.020.020.010.010.0020.0020.001\mathbf{0.001}
Real-time rate102%180%43%21%
ArtifactsYes ( 16(a) )Yes ( 16(b) )Yes ( 16(c) )No
TABLE III : Performance comparison between error-controlled integration (top) and fixed-step integration (bottom) on the dishrack simulation (Fig. 1 ). Error-controlled integration eliminates obvious artifacts (shown in Fig. 16 ) while achieving faster overall simulation speeds.
Fig. 17 : Snapshots of the dish rack task on hardware (top) and in simulation with CENIC (bottom, εacc=10−3\varepsilon_{\text{acc}}=10^{-3} ). The simulation and real trajectories are not exactly the same, due to model sensitivity and chaotic dynamics. But the overall qualitative behavior in simulation matches the real-world demonstration: there are no drops, missed grasps, passthrough incidents, rattling, or other major numerical artifacts. In contrast, standard discrete time-stepping methods introduce major artifacts, even with relatively small time steps (Fig. 16(a) - 16(c) ).
Fig. 18 : Real-time rate and step size δt\delta t for the complex dishrack simulation (Fig. 1 ) at accuracy εacc=10−3\varepsilon_{\text{acc}}=10^{-3} . By only taking small time steps when needed to resolve stiff collision dynamics, CENIC achieves faster overall simulation times than discrete time-stepping (Table III ).
Fig. 19 : Computational cost breakdown for each of the examples, run with error control at accuracy εacc=10−3\varepsilon_{\text{acc}}=10^{-3} .
Accuracy10−110^{-1}10−210^{-2}10−310^{-3}10−410^{-4}10−510^{-5}
Wall time (adaptive)0.190.390.651.112.69
Wall time (fixed)0.210.410.711.192.69
Iterations (adaptive)40159973192353953795839
Iterations (fixed)459110938208654072995839
TABLE IV : Impact of adaptive convergence criteria (Sec VI-B ) on hard clutter simulation times (in seconds) and total solver iterations.
Error controlReuseWall time (s)StepsFacts.Iters.
OffOff0.26100039803980
OffOn0.27100022936068
OnOff0.654991662416624
OnOn0.615101055719235
TABLE V : Impact of Hessian reuse on the hard clutter example. Hessian reuse reduces the number of expensive factorizations (“Facts.”), at the cost of more iterations. With error control enabled ( εacc=10−3\varepsilon_{\text{acc}}=10^{-3} , maximum timestep 0.1 s), this offers a small improvement in overall performance. In fixed step mode ( δt=10\delta t=10 ms), Hessian reuse does not improve performance.
Fig. 20 : Comparing different strategies for choosing an initial guess for linesearch, for the hard clutter problem at different accuracies. Our cubic spline method consistently reduces the number of iterations required for exact linesearch, resulting in modest improvements in overall speed.
Contact Stiffness (N/m)StepsIterationsWall time (s)
10210^{2}58522320.22
10410^{4}440111790.34
10610^{6}664247650.92
10810^{8}1378371831.742
101010^{10}1621346471.53
101210^{12}1641374951.63
TABLE VI : Statistics for hard clutter simulation (10 seconds of simulation time) at accuracy εacc=10−3\varepsilon_{\text{acc}}=10^{-3} , and varying contact stiffness kik_{i} . Stiffer contact parameters lead to more time steps, more iterations, and longer wall times. Nonetheless, CENIC is able to simulate this contact-rich scene significantly faster than real-time even with extremely stiff contact parameters (note that steel has stiffness around ki=107k_{i}=10^{7} N/m [ 17 ] ).
More in this topic

Robotics Simulation

Robotics simulation builds virtual robots and environments to model, train, and evaluate autonomous systems before real-world deployment.

See all 320 papers

Explore similar work

CardsList
Apr 11, 2026

GPU-Accelerated Continuous-Time Successive Convexification for Contact-Implicit Legged Locomotion

Contact-implicit trajectory optimization (CITO) enables the automatic discovery of contact sequences, but most methods rely on fine time discretization to capture all contact events accurately, which increases problem size and runtime while tying solution quality to grid resolution. We extend the recently proposed sequential convex programming (SCP) approach for trajectory optimization, continuous-time successive convexification (ct-SCvx), to CITO by introducing integral cross-complementarity constraints, which eliminate the risk of missing contact events between discretization nodes while preserving the flexibility of contact mode changes. The resulting framework, contact-implicit successive convexification (ci-SCvx), models full multibody dynamics in maximal coordinates, including stick-slip friction and partially elastic impacts. To handle complementarity constraints, we embed a backtracking homotopy scheme within SCP for reliable convergence. We implement this framework in a stand-alone Python software, leveraging JAX for GPU acceleration and a custom canonical-form parser for the convex subproblems of SCP to avoid the overhead of general-purpose modeling tools such as CVXPY. We demonstrate ci-SCvx on diverse legged-locomotion tasks. In particular, we validate the approach in MuJoCo with the Gymnasium HalfCheetah model against the MuJoCo MPC baseline, showing that a tracking simulation with the optimized torque profiles from ci-SCvx produces physically consistent trajectories with lesser energy consumption. We also show that the resulting software achieves faster solve times than existing state-of-the-art SCP implementations by over an order of magnitude, thereby demonstrating a practically important contribution to scalable real-time trajectory optimization.

Samuel C. Buckner, Purnanand Elango

Department of Aeronautics and Astronautics, University of Washington, Seattle, WA 98195. · Mitsubishi Electric Research Laboratories, Cambridge, MA 02139.

Trajectory OptimizationRobot LocomotionLegged RoboticsNonconvex Optimization
AskPDF
Sep 11, 2024

Highly-Efficient Differentiable Simulation for Robotics

Robotics simulators have improved significantly in computational speed and scalability, enabling them to generate years of simulated data for complex systems in minutes or hours. Despite these advances, efficiently and accurately computing simulation derivatives remains an open challenge. Addressing this would accelerate the convergence of reinforcement learning and trajectory optimization algorithms, particularly for contact-rich problems. This paper introduces a unifying framework for robotic simulation that accounts for all factors, including dynamics, collisions, and friction. The resulting algorithm computes analytical derivatives of the simulation by implicit differentiation, explicitly handling the intrinsic non-smoothness of the collision and frictional stages while exploiting the sparsity induced by the multi-body structure. Benchmark results demonstrate state-of-the-art performance, with timings ranging from 5 μ5\,μs for a 7-dof manipulator to 95 μ95\,μs for a 36-dof humanoid, an improvement of at least two orders of magnitude over alternative methods. Implemented in C++, the code will be open-sourced after the review process to support applications such as simulation-driven learning and real-time control.

Quentin Le Lidec, Louis Montaut, Yann de Mont-Marin +3

Inria, Ecole normale sup´erieure CNRS, PSL Research University 75005 Paris, France

Robotics SimulationMultibody DynamicsDifferentiable Physics
AskPDF
Oct 6, 2026

FlashNeRD: Performance-First Contact-Rich Neural Robot Dynamics

Compared with analytical physics, learned dynamics models promise robot simulation that is faster, inherently differentiable, and easily adaptable to real data. Neural Robot Dynamics (NeRD) pursues this by keeping collision detection analytical and replacing a simulator's numerical dynamics for the robot with a learned model. Three limitations remain. NeRD offers little speedup over the simulator it learned from, accepts contact only at predefined points, and has no two-way coupling with objects it manipulates. FlashNeRD removes all three with a parallel streaming architecture that makes each prediction faster and more accurate, an encoder that accepts contacts wherever they occur, and two-way coupling with objects simulated by analytical solvers. Experiments across five robots show faster and more accurate dynamics, faster policy learning, and faster inference-time planning. Across three robots, FlashNeRD's dynamics model is up to 55×55\times faster than an optimized NeRD and more accurate over long rollouts. This speedup extends to policy learning, where PPO trains an ANYmal locomotion policy in 34 seconds, 3.8×3.8\times faster than the analytical simulator and 2.7×2.7\times faster than an optimized NeRD. With DIAL-MPC, a sampling-based MPC method, the robot climbs all six test platforms where fixed-contact NeRD manages one, at approximately half the analytical simulator's planning time. Cube-reorientation policies trained with FlashNeRD complete within 2% of the simulator-trained policy's target count.

Mohammadmehdi Ataei, Mohammad Amin Nabian, Jie Xu +9

NVIDIA

Robot Dynamics LearningRobotics SimulationRobotic Control
AskPDF