TL;DR
This paper introduces a novel Learning Model Predictive Control (LMPC) strategy for autonomous racing that iteratively learns vehicle dynamics at the handling limit. By focusing on learning the error between a nominal model and actual vehicle behavior, their approach demonstrates improved robustness to parameter tuning and data scarcity, enabling safer and more efficient exploration of high-speed driving domains.
What Problem Does This Paper Address?
Autonomous racing, especially at the limits of vehicle handling and at high speeds, presents a significant challenge: to achieve higher speeds safely, you need accurate vehicle dynamics models, but to get accurate models, you need to drive at high speeds. This creates a "chicken-and-egg" problem. Existing Model Predictive Control (MPC) methods, while powerful, are fundamentally limited by the fidelity of their pre-defined vehicle models. Inaccurate models can lead to suboptimal performance, tracking errors, and, critically, dangerous overconfidence at the handling limit, potentially causing accidents. This paper aims to break this cycle by developing a control strategy that can safely learn and adapt to unknown vehicle dynamics in real-time, even with limited data.
Methods
The core innovation lies in modifying the Learning Model Predictive Control (LMPC) framework. Instead of trying to learn the entire vehicle dynamics model from scratch, this work proposes learning only the error dynamics. This means they start with a known, nominal physics-based model (which provides a safe baseline) and then learn a local, linear approximation of the difference between this nominal model's predictions and the vehicle's actual observed behavior.
Here's a breakdown of the key components:
- Nominal Model: A global, nonlinear, physics-based model of the vehicle is used as a starting point. This model is assumed to be reasonably accurate but not perfect, especially at the handling limit.
- Error Dynamics Regression: The crucial new part. The system models the true dynamics as $x^+ = f(x, u) + e(x, u)$, where $f$ is the nominal dynamics and $e$ is the unknown error. They then linearize this error term $e$ around a reference state and input. This linearized error dynamics is what they learn using data collected during operation.
- Local Affine Time-Varying (ATV) Models: The learned error dynamics are used to construct local affine models that approximate the true vehicle dynamics within specific operating regions.
- LMPC Framework: The learned ATV models are integrated into a standard LMPC framework. LMPC iteratively solves a finite-horizon optimal control problem (FHOCP) using data from previous laps to refine its control policy and adapt to changing dynamics. This includes synthesizing local convex target sets and cost-to-go functions from historical data.
The key difference from prior LMPC work is the focus on learning the error rather than the full dynamics, which is argued to improve robustness.
Datasets
The paper describes experiments using:
- Simulation Data: Used for initial testing and parameter tuning.
- 1/10th Scale Hardware Data: Experiments were conducted on a 1/10th scale autonomous race car platform (BARC).
- Full-Scale Autonomous Race Car Data: The strategy was deployed on a full-scale Indy Autonomous Challenge (IAC) race car at the Putnam Park Road Course. An initial dataset was collected using a tracking MPC at a relatively low speed (10 m/s) for three laps.
Key Insights and Ideas
- Learning the Error is More Robust: The central insight is that learning the difference between a known nominal model and the true dynamics is more robust to data scarcity and parameter tuning than trying to learn the entire dynamics from scratch. When data is sparse, the error learning approach naturally falls back to the more reliable nominal model, whereas full regression might produce inaccurate corrections.
- Safety through Nominal Model: The use of a nominal physics-based model provides an inherent safety fallback. Even if the learned error dynamics are imperfect, the controller still operates within the bounds of a known, albeit imperfect, physical model.
- Iterative Improvement at the Limit: The LMPC framework, combined with error dynamics learning, allows the vehicle to iteratively explore and improve its performance at the handling limit. This is visualized in Figure 1, showing how the trajectory adapts and speeds increase over laps.
- Breaking the Data Acquisition Cycle: The method directly addresses the challenge of acquiring high-speed data by enabling safe exploration and learning, thus breaking the cycle where good data requires high speeds, and high speeds require good data.
Key Results
- Improved Robustness: Experiments showed that the error dynamics regression LMPC is significantly more robust to parameter tuning (specifically, the bandwidth parameter 'h' and the control rate cost 'CRC') compared to prior full regression LMPC methods.
- Reduced Failures: In hardware experiments, particularly with low bandwidth settings (indicating data scarcity), the error regression LMPC exhibited fewer failures (e.g., leaving the track) compared to the full regression approach. For instance, in one hardware experiment, the error regression LMPC failed in only 1 out of 10 trials, while the full regression LMPC failed in 5 out of 10 trials.
- Comparable Performance: When parameters were well-tuned, the error regression LMPC achieved comparable lap times and learning performance to the full regression LMPC.
- Real-World Demonstration: The strategy was successfully demonstrated on a full-scale Indy Autonomous Challenge race car, showing iterative trajectory optimization and speed improvement around a hairpin turn.
Limitations and Future Work
- Nominal Model Accuracy: The robustness of the error regression approach is still dependent on the quality of the initial nominal model. While it can correct for mismatches, a severely inaccurate nominal model could still pose challenges. The authors suggest future work to examine the effect of the distance between the nominal and true system on stability.
- Generalization: While the paper focuses on local affine approximations, the generalization of these learned models to significantly different operating conditions or vehicle configurations is an area for further exploration.
- Multi-Vehicle Interaction: The current work focuses on single-vehicle racing. Extending this to multi-vehicle scenarios, which involve complex interactions and opponent prediction, is a natural next step.
Summary updated