Abstract
Researchers at Rutgers University have developed a hybrid approach to controlling low-cost, 3D-printed robots by combining analytical physics with machine learning. By using a "differentiable physics" simulator, they can automatically tune unknown friction parameters from just a few seconds of video data, enabling precise autonomous navigation where traditional black-box models fail.
Introduction and Problem Statement
High-end robots are typically easy to control because their physical properties—such as mass, inertia, and friction—are precisely calibrated during manufacturing. In contrast, "low-cost" robots built with 3D-printed parts and off-the-shelf components suffer from significant manufacturing uncertainties. Their wheels often have complex, irregular structures, and their friction against various floor surfaces is typically unknown and non-uniform.
Traditionally, bridging this "reality gap" requires either tedious manual tuning or pure data-driven methods like deep neural networks, which require massive amounts of training data. This research addresses the need for a data-efficient way to identify a robot's physical parameters and control it accurately in real-time.
Methodology
The authors propose a hybrid data-driven approach centered on a self-tuning differentiable physics simulator.
- Analytical Modeling: The team derived a mathematical model for the robot's Mecanum wheels, accounting for motor stall torque, angular velocity, and ground friction.
- Differentiable Physics: Unlike standard simulators, a differentiable simulator allows for the calculation of gradients—the direction and magnitude of change—of the error between simulation and reality.
- Optimization: Using these analytical gradients, the L-BFGS-B algorithm was employed to minimize a loss function based on the distance between the predicted and actual paths. This allows the system to "learn" specific friction coefficients for each wheel.
- Neural Network Integration: Recognizing that friction is dynamic and changes based on motor speed, a small neural network was trained to map control signals to the friction parameters identified by the physics model.
Experimental Setup and Results
The study utilized a custom-built $1200 mobile robot. Data was collected from eight unique trajectories, each lasting 6-8 seconds. Robot positions were tracked at 60fps using an overhead camera and Apriltags for ground truth. The proposed model was tested against standard derivative-free optimization methods (CMA-ES and Nelder-Mead) and pure data-driven neural networks.
Key Findings
- Data Efficiency: The method converges to an accurate model using only 40% of the collected data, making it viable for real-time, on-the-fly calibration.
- Speed and Accuracy: The gradient-based method was significantly faster and more reliable than Nelder-Mead (which failed to converge in 37% of cases) and CMA-ES.
- Path Following: The robot successfully followed complex "figure 8" curves with high precision. Conversely, a pure data-driven neural network trained on the same small dataset failed completely, with the robot quickly driving out of bounds.
Conclusion
By using physics as a prior, the system avoids the need to learn the laws of motion from scratch, focusing instead on identifying a few unknown parameters. While the current model assumes continuous motion and does not yet account for static friction or inertia at low power, it represents a significant step toward making low-cost autonomous hardware more capable and reliable. Future work will focus on incorporating these non-smooth physical effects into the differentiable framework.
Summary updated