Abstract
Localization is a fundamental challenge for autonomous robotics. Although accurate and efficient techniques now exist for solving this problem, they require explicit probabilistic models of the robot's motion and sensors. These models are usually obtained from time-consuming and error-prone measurement or tedious manual tuning. In this paper we examine automatic calibration of sensor and motion models from a Bayesian perspective. We introduce an efficient MCMC procedure for sampling from the posterior distribution of the model parameters. We also present a novel extension of particle filters to make use of our posterior parameter samples. Finally, we demonstrate our approach both in simulation and on a physical robot. Our results demonstrate effective inference of model parameters as well as a paradoxical result that using posterior parameter samples can produce more accurate position estimates than the true parameters.
Introduction
For an autonomous robot to navigate effectively, it must maintain an accurate estimate of its position (localization). This process relies on probabilistic models that describe how the robot moves (motion model) and how it perceives its environment (sensor model). Traditionally, these models are constructed through precise physical measurements or extensive manual tuning by experts. However, robots are dynamic: tires wear down, motors degrade, and environments change. Static models often fail to capture these shifts, leading to localization errors. This paper proposes a Bayesian framework to automatically calibrate these models from raw data, reducing human effort and improving robustness.
Methodology
The authors treat calibration as a Bayesian inference problem, seeking a posterior distribution of model parameters ($\theta$) given observed data ($D$). To handle the unknown true trajectory of the robot during training, they employ a hybrid Markov Chain Monte Carlo (MCMC) approach:
- Gibbs Sampling: Utilizing "Particle Smoothing" to sample likely robot trajectories based on current parameter estimates.
- Metropolis-Hastings: Proposing incremental changes to model parameters and accepting them based on how well they explain the sampled trajectories.
Once the parameter distribution is established, three methods for integration with Monte Carlo Localization (MCL) are evaluated: using the Mean of the samples, using the MAP (Maximum A Posteriori) estimate, or a fully Bayesian Posterior approach where the filter itself samples from the parameter distribution at each step.
Experimental Results
The framework was tested on a simulated Amigobot with sonar sensors and a physical Sony AIBO ERS-7 quadruped robot using vision-based landmark detection.
The Calibration Paradox
A key finding in the simulation studies was the "Calibration Paradox": localization using calibrated parameters often proved more accurate than using the "true" physical parameters that generated the data. As shown in Figure 3, the Mean and MAP calibration methods resulted in lower mean error than the ground-truth parameters. This occurs because the MCMC process finds parameters that compensate for the discretization errors inherent in the particle filter's finite approximation of the continuous world.
Real-World Validation
On the physical AIBO robot, the Bayesian calibration significantly outperformed the initial "Prior" model, which was based on expert manual tuning. The mean error was reduced from approximately 137mm to 125mm, a substantial improvement for a robot operating on a small soccer pitch. The results indicate that even short training sequences (e.g., 400 seconds) are sufficient to yield high-quality calibrations.
Conclusion
This work demonstrates that automatic Bayesian calibration is a powerful alternative to manual tuning. By optimizing parameters for the specific computational filter being used, the system can achieve higher accuracy than even physically "correct" models. While the current MCMC approach is a batch process, future research will focus on scaling to higher-dimensional parameter spaces and developing online calibration methods for real-time adaptation.
Summary updated