Chapter Four · failure evidence
What Kinematic & Dynamics Simulation got wrong, from 69 dissertations
The records describe various challenges encountered in kinematic and dynamic simulations, including sim-to-real transfer gaps, solver convergence breakdowns, and model inaccuracies. Researchers repeatedly found that simplifying assumptions such as linearity, rigid bodies, or pure kinematics failed to capture the complex contact, aeroelastic, and transient dynamics of physical systems. These records come from PhD theses at 17 institutions, 2021 to 2026. Each links to its thesis. They were extracted by language models reading the full text, so treat each as a lead to read, not a verdict.
Numerical solvers and optimization algorithms fail to converge on complex kinematic and dynamic equations
Formulations incorporating forward kinematics, leaf springs, or bounded non-smooth dynamics suffered from severe local minima, intractability, or mathematical infeasibility. In addition, high system dimensions and stiff dynamics triggered forward dynamics instability, model predictive control bound violations, and gradient variance blowup.
Tried and failed
quadratic programming with strict physiological bound constraints applied to musculoskeletal inverse dynamics torque estimation. Outcome: did not converge. Reason: kinematic irregularities and non-smooth dynamics made strictly bounded constraints mathematically infeasible
Understanding motor control and impairment: An upper-limb model for muscular assessment · EPFL
Tried and failed
trajectory optimization with compliant flexible elements applied to legged robot locomotion. Outcome: did not converge. Reason: modeling physical leaf springs in full rigid-body dynamics made solver convergence extremely difficult
Tried and failed
model-free reinforcement learning with reduced-order state applied to bipedal locomotion control. Outcome: did not converge. Reason: simplified low-dimensional state representation lacked sufficient dynamics information for whole-body motion synthesis
Towards versatile, high-performing and interactive humanoids · UT Austin
Tried and failed
program synthesis with forward kinematics constraints applied to manipulator inverse kinematics optimization. Outcome: did not converge. Reason: forward kinematics equations introduced numerous local minima, making numerical optimization intractable
Task-Based Design Synthesis Of Modular Manipulators · Cornell
Tried and failed
Restricting rhythmic pattern generation to proximal actuators applied to Neuromechanical forward dynamics gait generation. Outcome: did not converge. Reason: Failed to generate rhythmic locomotion when distal joint muscle pattern inputs were zeroed
Tried and failed
ADMM convex relaxation combining distance and velocity applied to joint kinematic parameter estimation. Outcome: too slow. Reason: excessive computational complexity and high processing overhead
Precise Geolocation for Drones, Metaverse Users, and Beyond: Exploring Ranging Techniques Spanning 40 KHz to 400 GHz · Virginia Tech
Tried and failed
projective dynamics differentiable simulation applied to low degree of freedom soft-body dynamics. Outcome: worse than baseline. Reason: prefactorization overhead provides no advantage over exact Newton steps at small system dimensions
Differentiable Simulation Methods for Robotic Agent Design · MIT
Considered and rejected
Considered and rejected: Rejected solving Hamilton-Jacobi-Bellman PDEs for reachability due to intractable high dimensionality of multi-body humanoid dynamics.
Whole-body trajectory generation and control strategies for multi-contact robots · UT Austin
Considered and rejected
Considered and rejected: Rejected using OpenSim directly for large-scale RL due to extreme computational slowness and forward-dynamics instability.
Reinforcement Learning for Muscle-Driven Systems · Publikationssystem UB Tuebingen
Considered and rejected
Considered and rejected: Rejected computing control inputs via continuous analytical inverse dynamics because redundant, non-linear activation dynamics and ODE-based muscle contraction models prohibit direct mathematical inversion.
Neuro-musculoskeletal Models: A Tool to Study the Contribution of Muscle Dynamics to Biological Motor Control · Publikationssystem UB Tuebingen
Considered and rejected
Considered and rejected: Searching directly over joint angles theta using forward kinematics equations (rejected due to highly non-linear functions creating numerous local minima; replaced by parameterizing states P directly)
Task-Based Design Synthesis Of Modular Manipulators · Cornell
Tried and failed
model predictive control with learned neural kinematics applied to trajectory tracking in flexible robot manipulation. Outcome: unstable. Reason: model inaccuracies and unhandled uncertainty caused critical bound violations and tracking failure
Robot model learning and optimal control towards safer robotic surgery · Imperial
Tried and failed
gradient estimation through differentiable physics simulation applied to stiff or discontinuous dynamical systems. Outcome: unstable. Reason: gradients suffer from empirical bias, zero variance illusion, and severe variance blowup in discontinuous dynamics
Leveraging Structure for Efficient and Dexterous Contact-Rich Manipulation · MIT
Rigid body and isolated component models neglect structural flexibility and aeroelastic interactions
Assuming rigid bodies or analyzing components in isolation overlooked critical physical behaviors such as material springback, structural deflection, and mutual aerodynamic interactions. These omissions led to missed vibration peaks, underpredicted roll acceleration, physically infeasible jumping trajectories, and poor aerodynamic torque estimations.
Tried and failed
contact-rich rigid body dynamics simulation applied to mechanical adhesion in swarm robotics. Reason: Microscopic mechanical interlocking could not be simulated natively, requiring artificial constraint joints to match physical adhesion.
Collision Induced Self Organization in Shape Changing Robots · Georgia Tech
Tried and failed
musculoskeletal simulation omitting muscle force-length-velocity dynamics applied to biomechanical gait simulation. Outcome: unstable. Reason: neglecting dynamic muscle properties caused reserve actuator forces to exceed acceptable validity thresholds
Predicting lower limb kinematics and kinetics from internal measurement units using deep learning · Imperial
Tried and failed
centroidal dynamics with full kinematics trajectory optimization applied to humanoid dynamic jumping maneuvers. Reason: omits joint torque and electrical power limits, yielding physically infeasible trajectories
A Model-Based Planning and Control Framework for Parkour-Style Legged Locomotion · MIT
Tried and failed
free-floating rigid body dynamics modeling applied to spacecraft attitude during in-situ manufacturing. Outcome: did not generalise. Reason: neglected metal springback effects caused unmodeled rotational steps and vibrations
Rapid In-Space Assembly and Manufacturing of Large Reticulated Truss Structures · MIT
Tried and failed
rigid-body equations of motion applied to flexible high aspect ratio structures. Reason: over-damped dynamics caused underprediction of roll acceleration and missed peak oscillations
Handling qualities of high aspect ratio wing aircraft. · Cranfield
Tried and failed
inertial-only multibody dynamics modeling applied to aerodynamic pitch angle prediction. Outcome: worse than baseline. Reason: aerodynamic moments acted in opposition to inertial forces, which were neglected
Multibody Dynamics of Aeromechanical Systems · Virginia Tech
Tried and failed
articulating nose for direct roll control applied to guided projectile attitude stabilization. Reason: out-of-plane deflections generate weak roll moments that are strongly coupled to longitudinal dynamics
Smart Projectile Performance Augmentation Using an Articulating Nose · Georgia Tech
Tried and failed
static wind tunnel testing for drag estimation applied to multirotor aerodynamic calibration. Outcome: did not generalise. Reason: fixed rotors neglect dynamic rotor aerodynamics and induced airflow present during flight
Spatially distributed Wind and Turbulence Measurements with a Fleet of Unmanned Aerial Systems · Publikationssystem UB Tuebingen
Tried and failed
coupling distinct unilateral structural models applied to biological wing structure modeling. Outcome: did not generalise. Reason: high kinematic indeterminacy preventing accurate structural thickness predictions
Geometry And Topology: Building Machine Learning Surrogate Models With Graphic Statics Method · Penn
Tried and failed
neural ODE residual learning without domain physics applied to aerodynamic torque and proximity dynamics. Outcome: did not generalise. Reason: omitting nominal downwash aerodynamic models prevented accurate out-of-distribution torque predictions in close-proximity conditions
Learning-based Model Predictive Control for Aerial Vehicles · Penn
Tried and failed
single-component proxy modeling with partial loading applied to multi-blade rotor aeroelastic kinematics. Outcome: did not generalise. Reason: isolated blade aerodynamic trimming failed to reproduce multi-blade aerodynamic interactions and structural deformations
Computational Investigation of Separated Flow and Stall Events on Rotating Systems · Georgia Tech
Lost to a baseline
Beam model overpredicted baseline wing torsional deformation compared to shell FEM due to the kinematic omission of rib stiffness in the beam formulation
Linear approximations and steady state models fail to capture complex nonlinear transient dynamics
Linearized state space representations, linear regressions, and steady state analyses failed when applied to nonlinear multibody systems, vehicle handling, and maneuvering targets. These simplified linear models could not capture actuator saturation, nonlinear steering geometry, or transient spatial temporal behaviors.
Tried and failed
PID tuning based on simplified linear plant models applied to high-fidelity nonlinear vehicle dynamics control. Outcome: did not generalise. Reason: linearised single-track model transfer functions failed to capture higher-order nonlinear dynamics of high-fidelity simulation
Optimal handling characteristics for electric vehicles with torque vectoring. · Cranfield
Tried and failed
linearization to state-space models applied to nonlinear multibody transient dynamics. Outcome: did not generalise. Reason: produced inaccurate transient responses and was computationally inefficient across varying operating points
A Multi-fidelity Modeling Approach for Vehicle Performance Tuning · Georgia Tech
Tried and failed
extended Kalman filtering with kinematic motion models applied to trajectory prediction of maneuvering targets. Outcome: worse than baseline. Reason: linear kinematic assumptions cannot accurately capture complex nonlinear maneuvering dynamics
Data-driven Target Tracking and Hybrid Path Planning Methods for Autonomous Operation of UAV · Virginia Tech
Tried and failed
Multiple linear regression baseline applied to Traffic incident delay estimation. Outcome: worse than baseline. Reason: Linear assumptions cannot capture complex nonlinear spatial-temporal traffic dynamics
Tried and failed
linearised reduced-order structural dynamics modelling applied to flexible wing aeroservoelastic control. Reason: linear models failed to accurately capture in-plane structural dynamics compared to out-of-plane modes
Tried and failed
steady-state dynamic response analysis applied to tracked vehicle ride vibration modeling. Outcome: did not generalise. Reason: steady-state analysis systematically overestimated vibration levels compared to true transient dynamics on sinusoidal terrain profiles
Tried and failed
linear regression metamodeling of simulation outputs applied to manufacturing assembly time surrogate modeling. Outcome: did not generalise. Reason: linear models could not capture non-linear simulation dynamics, consistently underestimating completion times
Tried and failed
linear system identification of inverse dynamics applied to vehicle braking systems. Outcome: did not generalise. Reason: linear models failed to capture non-linearities like ABS activation and traction saturation
Adaptive Longitudinal and Lateral Control for Autonomous Vehicles: High-Speed Platooning of Articulated Trucks · Virginia Tech
Tried and failed
constant steering ratio in kinematic model applied to vehicle trajectory prediction. Outcome: did not generalise. Reason: nonlinear steering geometry causes large trajectory divergence under moderate and rapid steering inputs
Parameter Identification and Validation of a Control-Oriented Vehicle Dynamics Model for an Autonomous Chevrolet Bolt EUV · Virginia Tech
Considered and rejected
Considered and rejected: Rejected model-based RL and conventional convex optimization control policies due to large biases from oversimplification of complex nonlinear microscopic traffic dynamics.
Modelling urban street configurations for Autonomous Vehicle flows · Imperial
Pure kinematic representations fail when neglecting forces, compliance, and contact friction
Simulating mechanisms through pure kinematics or prescribed displacements failed because friction, compliance, and contact loads were ignored. These deficiencies led to unmodeled toolpath discrepancies, paradoxical joint sliding, trajectory tracking errors, and foot slip during locomotion.
Tried and failed
multi-axis force-controlled robotic kinematic simulation applied to isolated knee replacement implant kinematics. Reason: lack of soft-tissue constraints caused paradoxical anterior femoral sliding under posterior drawer loads
Robotic testing of total knee replacement designs · Imperial
Tried and failed
discrete event simulation for robot kinematics estimation applied to robotic process time estimation. Reason: discrete-event modeling lacked the physics fidelity required to accurately compute robotic movement and cycle times
Considered and rejected
Considered and rejected: Kinematic draping simulations (KDS) were rejected for multi-ply forming analysis due to their fundamental inability to model friction, compaction, or out-of-plane bending physics.
Automated forming of multi-ply non-crimp fabric structures · University of Nottingham Repository
Tried and failed
heuristic trajectory generation decoupling dynamics from planning applied to bipedal locomotion over large obstacles. Outcome: unstable. Reason: heuristic swing trajectories ignored full-body dynamics, causing foot slip on high steps
Tried and failed
kinematics-based series elastic actuator force modeling applied to robotic exoskeleton force estimation. Reason: unmodeled friction and compliance/deformations in mechanical linkages caused high force prediction errors
Vision-Based Force Planning and Voice-Based Human-Machine Interface of an Assistive Robotic Exoskeleton Glove for Brachial Plexus Injuries · Virginia Tech
Tried and failed
prescribed displacement boundary condition in FE modeling applied to unconstrained frictional sliding dynamics. Outcome: did not generalise. Reason: prescribed kinematics artificially constrained physical contact dynamics and sliding interface behavior
Variable frictional interface behaviour for nonlinear dynamic response control · Imperial
Tried and failed
continuum kinematic modeling with idealized contact mechanics applied to locomotion speed prediction of articulated robots. Outcome: did not generalise. Reason: rigid link interiors caused unintended sliding drag instead of idealized pure rolling ground contact
Shape-centric Modeling for Control of Snake-like Robots · Georgia Tech
Considered and rejected
Considered and rejected: Rejected analytical kinematic fishnet/mapping models because they neglect forming loads, frictional interactions, and bending stiffness necessary for defect prediction.
Thermomechanical forming simulation for fibre reinforced thermoplastic laminates · University of Nottingham Repository
Considered and rejected
Considered and rejected: Rejected purely kinematic TP approaches because neglecting external forces (thrust, drag, wind) severely degrades prediction accuracy across dynamic flight phases
Trajectory modelling and execution for multi-Unmanned Aerial Vehicle applications · Imperial
Policies and open loop trajectories fail during sim to real transfer due to unmodeled contact and friction
Trajectories and reinforcement learning policies optimized within physics engines failed to generalize when deployed directly onto physical hardware. The sim to real gap arose because simulators lacked accurate physical dynamics, real time adaptation, and interactive contact friction models.
Tried and failed
open-loop transfer of simulation-optimized trajectories to hardware applied to locomotion controller for physical robots. Outcome: did not generalise. Reason: sim-to-real gap and unmodeled contact dynamics caused unpredictable execution without feedback
Robot Graph Grammars: Towards Custom Robots for Every Task · MIT
Tried and failed
direct sim-to-real transfer of reinforcement learning policies applied to embodied robotic navigation. Outcome: did not generalise. Reason: unmodeled real-world physical dynamics mismatch between the simulation environment and physical hardware
4D audio-visual learning: a visual perspective of sound propagation and production · UT Austin
Tried and failed
direct sim-to-real transfer of synthesized control policies applied to real-world robotic control. Outcome: did not generalise. Reason: simulated dynamics did not match real-world physical interactions without real-time adaptation
Leveraging program synthesis for robust long-term robot autonomy via interactive learning and adaptation · UT Austin
Tried and failed
open-loop trajectory tracking applied to contact-rich robotic assembly insertions. Outcome: did not generalise. Reason: unmodeled contact dynamics and real-world execution errors caused alignment failures without feedback
Considered and rejected
Considered and rejected: Rejected end-to-end DRL learning of guidance and low-level control directly from states to actuator forces/torques because policy overfits simulated dynamics and fails during sim-to-reality transfer.
Deep Reinforcement Learning as Guidance for Aerospace Robotics · Carleton University Institutional Repository
Considered and rejected
Considered and rejected: Rejected direct velocity-based state observations in favor of position-based observations due to difficulties in accurately modeling robot dynamics during sim-to-real transfer.
Tried and failed
training regression models on frictionless simulation data applied to mechanical stiffness estimation. Outcome: did not generalise. Reason: frictionless simulation failed to capture real-world friction dynamics leading to catastrophic evaluation errors
ULTRASONIC IMAGING AND TACTILE SENSING FOR ROBOTIC SYSTEMS · Georgia Tech
Tried and failed
Direct open-loop execution of extracted trajectory priors applied to robotic grasping and manipulation tasks. Outcome: did not generalise. Reason: Demonstration trajectories without interactive exploration cannot handle real-world physical and kinematic discrepancies.
Teaching Robots using Interactive Imitation Learning · Virginia Tech
Kinematic estimators and predictions lose performance compared to dynamic models and simpler baselines
Kinematics only predictors and uncorrected kinematic formulations performed worse than classical baselines and dynamic model formulations. These approaches failed joint limit constraints, yielded inferior motion predictions for stroke survivors, and produced large trajectory errors.
Lost to a baseline
Differential inverse kinematics achieved lower RMSE (0.0036m x, 0.0044m y, 0.00055m z) than gradient projection (0.0037m, 0.0034m, 0.0070m) and ANFIS (0.1309m, 0.0209m, 0.1261m), but failed joint limit constraints.
Development of Methodologies, Mechatronic Solutions and Controls for Upper Body Rehabilitative Robotics · IRIS - POLITO - prod
Lost to a baseline
Forward kinematics with 6 base DoFs yielded higher positional error than inverse kinematics optimization using the same 6 DoFs
An integrative computational modeling approach for Drosophila motor control · EPFL
Lost to a baseline
Kinematic prediction by interpolating gait dynamics was worse than direct kinematic interpolation for two low-functioning stroke survivors who deviated furthest from the able-bodied centroid.
Gait signatures: data-driven discovery of individual-specific neuromechanical dynamics · Georgia Tech
Lost to a baseline
MPL gait model yielded poor torque correlations compared to human data (rThip = 0.57, rTknee = 0.46, rTank = 0.47) and worse ankle kinematics than Geyer & Herr (2010) (rankle = 0.42).
Modular Neural Controllers for Predictive Simulations of Lower Limb Tasks · Research Repository UCD
Lost to a baseline
Self-driven prediction of stroke kinematics from initial posture performed worse than measured stride-to-stride variation (60% of other gait cycles were closer to the target than the self-driven prediction, compared to only 21% in able-bodied individuals).
Gait signatures: data-driven discovery of individual-specific neuromechanical dynamics · Georgia Tech
Lost to a baseline
In microscopic traffic simulation on I24-MSD, standard SMART (CE) had worse kinematic score (0.7353) than classical IDM (0.7592) and Constant Speed (0.7581) baselines
Learning to Tackle Task Variations in Control - A Transportation Context · MIT
Lost to a baseline
In autonomous racing, BestCase MPC with full dynamic model knowledge beat BayesRace GP-corrected kinematic model prior to multi-lap model updates
Lost to a baseline
Kinematical approach with orbit error (Case 4) was 10 to 100 times worse than zero-orbit-error case (Case 3c) and failed to match simulated dynamical GRACE performance
Euler angle parameterizations suffer from kinematic singularities during dynamic rotations
Using Euler angle representations and base frame kinematic models introduced algorithmic and kinematic singularities during dynamic motions and aggressive maneuvers. These parameterizations produced representation ambiguities and poor metric distances, prompting the adoption of coordinate free representations on Lie groups.
Considered and rejected
Considered and rejected: Rejected Euler angles (Tait-Bryan) for spacecraft rotational dynamics representation due to representation ambiguity and kinematic singularities at pitch theta = pi/2.
Vision based Real-Time Navigation with Unknown and Uncooperative Space Target · Carleton University Institutional Repository
Considered and rejected
Considered and rejected: Rejected standard Euler-angle rotational dynamics parameterization in SRB MPC in favor of Variational Based Linearization on SO(3) to avoid kinematic singularities during dynamic motions
Real-Time Planning and Nonlinear Control for Robust Quadrupedal Locomotion with Tails · Virginia Tech
Considered and rejected
Considered and rejected: Rejected standard Euler-angle orientation error metrics for aggressive maneuvers because of kinematic singularities and poor metric distance at large angles.
Considered and rejected
Considered and rejected: Rejected base-frame-based RCM kinematic formulations due to algorithmic singularities occurring independent of mechanical configuration
Development and Evaluation of Control Methods with Dynamic RCM Constraint on Surgical Robotics System in Minimally Invasive Surgery · JScholarship
Left open by the authors
Problems the authors named and did not get to.
Left open
Estimate individual tendon tensions and build online-varying backlash and hysteresis models for tendon-driven surgical robots. Blocker: Requires the physical Micro-IGES surgical robot hardware or dedicated experimental apparatus to measure and validate tendon tension dynamics.
Robot model learning and optimal control towards safer robotic surgery · Imperial
Left open
Develop an optimal weight tuning method for the MPC cost function in surgical tendon-driven robot control. Blocker: Requires the physical Micro-IGES surgical robot setup or its proprietary kinematic simulation data
Robot model learning and optimal control towards safer robotic surgery · Imperial
Left open
Extend cross-learning to navigation problems featuring complex dynamics from human-machine interaction. Blocker: Vague task direction lacking specific formulation, targets, or interactive dynamics models
Left open
Compute lower-limb joint quasi-stiffness from perturbation walking kinematics and kinetics data and compare it to steady-state gait. Blocker: Requires experimental biomechanical gait perturbation data (motion capture and force plates) or specialized perturbation simulations.
An assessment of lower-limb joint quasi-stiffness during healthy and transtibial amputee walking · UT Austin
Left open
Train a machine learning model on human kinematic data to generate nominal crawling trajectories parameterized by worker body dimensions. Blocker: Requires a comprehensive motion-capture dataset of diverse human subjects crawling.
Analysis, Design, and Control of Supernumerary Robotic Limbs Coupled to a Human · MIT
Left open
Extend Complex Wrinkle Field kinematics to simulate dynamic cloth PDEs beyond boundary value interpolation. Blocker: Lacks specific mathematical formulations or numerical algorithms to couple full dynamic cloth equations with CWF kinematics
Complex wrinkle simulation and robust surface remeshing · UT Austin
Left open
Extend the optimal local truncation error method to simulate stationary and propagating cracks and non-linear partial differential equations. Blocker: High mathematical complexity with no specified formulation for non-linearities or crack propagation dynamics
Left open
Train machine learning models (Random Forest, SVM, Deep Learning) on LiDAR trajectory data to classify driver behavior instead of rule-based kinematic thresholds. Blocker: Access to the thesis author's roadside LiDAR trajectory dataset and proprietary VISSIM simulation environment.
Calibration of Microscopic Traffic Simulation Models for Proactive Safety Performance based on LiDAR Trajectory Data · Texas Tech
Left open
Perform scale-resolving simulations to model unsteady aerodynamics for active aerodynamic control of high-performance vehicles at extremely low ride heights. Blocker: Lack of specific proprietary vehicle geometry and boundary conditions from the original thesis
Analysis of active aerodynamics for high-performance vehicles · Cranfield
Left open
Develop and test model-based predictive control and autonomy algorithms within the multibody dynamics simulation framework. Blocker: Access to the thesis author's specific Simscape Multibody model files and vehicle parameters
Multibody dynamics modeling, validation, and analysis of a mobile off-road vehicle · Iowa State
Checking a claim in this area?
We can run the same search on any method or claim. If nothing turns up, we will say so, and that proves nothing on its own.