Integration of the TIAGo Robot into Isaac Sim with Mecanum Drive Modeling and Learned S-Curve Velocity Profiles

Basic Information

• Paper ID: 2510.10273
• Title: Integration of the TIAGo Robot into Isaac Sim with Mecanum Drive Modeling and Learned S-Curve Velocity Profiles
• Authors: Vincent Schoenbach¹, Marvin Wiedemann², Raphael Memmesheimer¹, Malte Mosbach¹, Sven Behnke¹
• Institutions: ¹University of Bonn (Autonomous Intelligent Systems), ²Fraunhofer Institute for Material Flow and Logistics
• Classification: cs.RO (Robotics)
• Publication Date: October 14, 2025 (arXiv v2)
• Paper Link: https://arxiv.org/abs/2510.10273
• Code Link: https://github.com/AIS-Bonn/tiago_isaac

Abstract

This paper presents the integration of the PAL Robotics TIAGo++ Omni mobile manipulator into the NVIDIA Isaac Sim simulation environment. The robot is equipped with a Mecanum wheel chassis enabling omnidirectional mobility. The authors propose two control models: a physically accurate model and a lightweight velocity model, and employ learning methods to calibrate the robot's S-curve velocity profiles. This simulation model provides researchers with an efficient platform for conducting learning-based control experiments in diverse environments.

Research Background and Motivation

Problem Definition

1. Core Challenge: The absence of an available simulation model for the TIAGo++ Omni robot in Isaac Sim limits researchers' ability to leverage modern GPU-accelerated simulation environments
2. Technical Difficulties:
   • Gazebo models are not directly compatible with Isaac Sim
   • Omnidirectional drive via Mecanum wheels requires precise modeling involving computationally intensive physics calculations
   • Lack of access to internal controllers makes it difficult to reproduce real wheel acceleration behavior
   • Traditional PID controller tuning cannot accurately reflect real robot behavior

Research Significance

1. Simulation Advantages: Modern physics simulation has become foundational to robotics research, providing cost-effective and scalable training and evaluation environments
2. GPU Acceleration Requirements: GPU-accelerated frameworks like Isaac Sim are particularly suited for learning-based methods, enabling generation of large-scale training data
3. Safety Considerations: Simulation provides safe training and testing environments, minimizing risk of damage to robots and surroundings

Limitations of Existing Approaches

1. Gazebo Limitations: Limited physical accuracy and graphical realism, making it difficult to meet requirements for precise modeling and environmental perception
2. Compatibility Issues: Existing robot models are primarily designed for Gazebo, making integration with modern machine learning pipelines challenging
3. Omnidirectional Drive Modeling: Most research employs approximate methods rather than physically accurate Mecanum wheel simulation

Core Contributions

1. Physically Accurate Model: Provides high-fidelity simulation of TIAGo++ Omni, including precise physical modeling of Mecanum wheels
2. Lightweight Control Model: Develops computationally efficient velocity control model suitable for large-scale learning applications
3. Neural Network Calibration Method: Approximates real robot S-curve velocity profiles using learning methods with minimal trajectory data recording
4. Open-Source Integration: Provides complete Isaac Sim extension supporting ROS 2 integration and sensor simulation
5. Design Insights: Reveals the importance of proportional acceleration in Mecanum wheel robot controller design

Methodology Details

Task Definition

Complete integration of the TIAGo++ Omni robot into Isaac Sim, with emphasis on solving omnidirectional drive simulation. Input consists of Twist commands (linear velocities vx, vy and angular velocity vθ), with output being accurate robot motion behavior.

Model Architecture

1. Physically Accurate Simulation Model

Mecanum Wheel Modeling:

• Procedurally generated Mecanum wheels containing 15 freely rotating rollers at 45-degree angles
• Each roller modeled using 6 spheres as collision bodies, balancing computational efficiency and physical accuracy
• Ensures precise roller alignment and wheel geometry

Kinematic Control: The relationship between wheel speeds and chassis velocity is represented by the following matrix:

$\begin{bmatrix} \omega_1 \\ \omega_2 \\ \omega_3 \\ \omega_4 \end{bmatrix} = \frac{1}{r} \begin{bmatrix} 1 & 1 & -(L_x + L_y) \\ 1 & -1 & (L_x + L_y) \\ 1 & -1 & -(L_x + L_y) \\ 1 & 1 & (L_x + L_y) \end{bmatrix} \begin{bmatrix} v_x \\ v_y \\ v_\theta \end{bmatrix}$

where r is the wheel radius, and Lx and Ly are the distances from wheels to the robot center.

2. Lightweight Velocity Model

• Directly sets robot chassis velocity, avoiding wheel-ground interaction calculations
• Still applies S-curve velocity profiles to maintain smooth acceleration
• Significantly improves simulation efficiency, reducing physics steps from 360 to 60 steps/second

3. S-Curve Velocity Profile Learning

Neural Network Architecture:

• Small MLP: 1, 35, 15, 5 layer structure
• Softplus activation functions
• Outputs 5 parameters: Θω = (a, b, m, k1, k2)

S-Curve Model: Piecewise function composed of three segments:

1. Initial phase: smooth rising function s1 (interval 0,a)
2. Linear transition: linear function s2 (interval a,b)
3. Final phase: smooth asymptotic function s3 (interval [b,∞))

Velocity Interpolation Strategy: Transition from current velocity T to target velocity T': $T_t = T + p_t \cdot (T' - T)$ where $p_t = \frac{S_\Theta(t-t_0)}{\Delta\omega}$ represents the proportion of command executed at time t.

Technical Innovations

1. Data-Driven Calibration: Uses neural networks to learn real robot acceleration characteristics rather than manual parameter tuning
2. Piecewise S-Curve Design: Customized three-segment S-curve model that better matches real data than standard logistic functions
3. Dual-Model Architecture: Provides accuracy-efficiency trade-off choices for different application scenarios
4. Minimal Data Requirements: Generalizes to complex motions using only basic directional motion trajectory data

Experimental Setup

Data Collection

• Motion Types: Basic trajectories in x-direction, y-direction, and rotational motion
• Velocity Range: Linear velocities 0.05-1.00 m/s, angular velocities 0.05-1.5 rad/s
• Data Processing: Each command repeated 3 times with averaging to reduce noise
• Sensors: Motion capture using OpenVR and VIVE trackers

Evaluation Metrics

Mean Relative Error Definition: $\text{MRE} = \frac{|\Delta p_{\text{real}} - \Delta p_{\text{sim}}|}{\Delta p_{\text{real}}}$

where Δp represents total travel distance (linear motion) or total rotation angle (rotational motion).

Test Trajectories

1. Square Trajectory: forward → right → backward → left, 3 seconds per segment, target velocity 0.45 m/s
2. Circular Trajectory: y-direction 0.19 m/s + z-axis rotation 0.78 rad/s
3. Single-Axis Motion: 20 different target velocities in x, y, and diagonal directions

Experimental Results

Main Results

Key Findings

1. Lightweight Model Performance: Comparable to physically accurate model in most cases while achieving significantly improved computational efficiency
2. Generalization Capability: Model generalizes to diagonal motions not seen during training
3. Controller Limitations: Reveals that PAL Robotics' original controller does not follow proportional acceleration principles in certain compound motions
4. Trajectory Accuracy: High simulation accuracy for simple trajectories with cumulative errors in complex trajectories

Case Analysis

• Square Trajectory: Simulation produces smoother turns than real robot, which actually increases control difficulty and may improve robustness
• Circular Trajectory: Overall trajectory shape is well-preserved with minor position offsets

Related Work

Omnidirectional Wheel Modeling Research

• Most research employs traditional simulation tools such as Matlab-Simulink, RecurDyn, and Gazebo
• Few works explore omnidirectional robot simulation in Isaac Sim
• Existing research primarily focuses on kinematic modeling and motion verification

Real2Sim Methods

• Traditional approaches rely on manual model tuning
• Machine learning methods for system identification but difficult to integrate with non-differentiable physics engines
• Optimization methods such as evolutionary algorithms for simulation parameter tuning

Uniqueness of This Work's Contributions

• First neural network modeling of S-curve velocity profiles for Mecanum wheels
• Focuses on open-loop velocity models rather than closed-loop trajectory tracking
• Seamless integration with Isaac Sim API

Conclusions and Discussion

Main Conclusions

1. Successfully integrated TIAGo++ Omni into Isaac Sim with two control model options
2. Learning methods effectively approximate real robot S-curve velocity profiles
3. Lightweight model significantly improves computational efficiency while maintaining reasonable accuracy
4. Proportional acceleration is crucial for smooth omnidirectional motion in Mecanum wheel robots

Limitations

1. Non-Proportional Acceleration: Design flaws in the original controller make certain motion patterns difficult to model accurately
2. Open-Loop Control: Only models open-loop behavior; practical applications require closed-loop controllers
3. Data Dependency: Model accuracy limited by quality and coverage of training data
4. Cumulative Error: Error accumulation occurs in long-duration complex trajectories

Future Directions

1. Further validate framework performance in reinforcement learning and trajectory tracking tasks
2. Explore alternative methods for improving S-curve velocity profile modeling in Mecanum wheel robots
3. Extend to other omnidirectional robot platforms
4. Develop more accurate compound motion modeling methods

In-Depth Evaluation

Strengths

1. High Practical Value: Fills the gap of dual-arm omnidirectional robot models in Isaac Sim
2. Methodological Innovation: Cleverly combines data-driven methods with physics simulation to solve black-box controller modeling challenges
3. Engineering Completeness: Provides complete open-source solution including ROS 2 integration and sensor support
4. Performance Trade-off: Dual-model design offers accuracy-efficiency trade-offs for different application scenarios
5. In-Depth Analysis: Reveals important principles in Mecanum wheel controller design

Limitations

1. Modeling Constraints: Limited accuracy in certain motion patterns due to non-ideal original controller design
2. Data Coverage: Relatively simple training data; generalization to complex motions requires further verification
3. Validation Scope: Primarily open-loop validation; lacks assessment in closed-loop control and learning tasks
4. Theoretical Analysis: Lacks theoretical guidance for S-curve parameter selection and network architecture design

Impact

1. Academic Contribution: Provides important tools for robotics simulation community, promoting TIAGo-related research
2. Practical Value: Lowers research barriers, enabling more researchers to utilize modern simulation environments
3. Methodological Inspiration: Provides reference framework for Isaac Sim integration of other robot platforms
4. Open-Source Impact: Public code and models promote community collaboration and development

Applicable Scenarios

1. Reinforcement Learning: Large-scale parallel training, particularly suited for lightweight models
2. Perception Research: Leveraging Isaac Sim's high-quality rendering for visual algorithm development
3. Navigation Algorithms: Testing omnidirectional movement strategies in complex environments
4. Human-Robot Interaction: Research in dual-arm collaboration and mobile manipulation tasks
5. Algorithm Verification: Safe testing before deployment to real robots

References

This paper cites 22 relevant references covering multiple domains including robot simulation, Mecanum wheel modeling, and system identification, providing solid theoretical foundation and technical background for the research.