Model-Based Task Transfer Learning for Iterative Learning Controllers
A model-based task transfer learning (MBTTL) method is presented. We consider a constrained nonlinear dynamical system and assume the availability of state-input pair datasets which solve a task T1. Our objective is to find a feasible state-feedback policy for a second task, T2, by using stored data from T1. Our approach applies to tasks T2 which are composed of the same subtasks as T1, but in different order.
In this paper we formally introduce the definition of subtask, the MBTTL problem and provide examples of MBTTL in the fields of autonomous cars and manipulators. Then, a computationally efficient approach to solve the MBTTL problem is presented along with proofs of feasibility for constrained linear dynamical systems. Simulation results show the effectiveness of the proposed method.
Applying the MBTL algorithm to an ILC guarantees persistent feasibility and reduces the task iteration cost compared with planning from scratch. Furthermore, The effectiveness of the MBTL algorithm improves with each continued application. After three rounds, the MBTTL-initialization results in a $26\%$ reduction in iteration cost over five iterations compared to PID-initialization:
C. Vallon and F. Borrelli: “Model-Based Task Transfer” submitted to 2019 IEEE Conference on Decision and Control (CDC). Get Paper
Reachability for Hybrid Path Planning
Path planning for autonomous vehicles must often be done in real-time in order to ensure safety and robustness of the control algorithm. Real-time planning using optimal control algorithms can have prohibitively high computational cost, particularly if high-fidelity vehicle models are used in the prediction steps. This motivated the design of the previously published FaSTrack algorithm, which uses a simple planner model along with a conservative error bound as an alternative to restrictively complex vehicle models. We propose an extension to the FaSTrack algorithm that results in offline classification of race track maneuverability. We perform forward reachability analyses at different minimum velocities in order to divide a race track into discrete modes which represent minimum model fidelity requirements for online path planners as a function of desired vehicle velocity.
This segmentation can then be used as safety constraints during real-time path planning along that track or to determine a priori if a particular path planning algorithm will be appropriate. We demonstrate our algorithm on a simple example with two planner models and find that the resulting segmentation is indeed representative of track section complexity.
Sections that require the higher-fidelity planner model are marked using a dotted blue line. Only the most dangerous areas are highlighted at lower velocities, while increased speeds lead to the identification of more of the track as a potential hazard.
Personalized behavior models for safe lane change initiation and control
C. Vallon, Z. Ercan, A. Carvalho and F. Borrelli: “A machine learning approach for personalized autonomous lane change initiation and control” in 2017 IEEE Intelligent Vehicles Conference (IV). PDF
Lane change initiation in autonomous driving is typically based on subjective rules, functions of the positions and relative velocities of surrounding vehicles. This approach is often arbitrary, and not easily adapted to the driving style preferences of an individual driver. Here we propose a data-driven modeling approach to capture the unique lane change decision behavior of human drivers. We collect data with a test vehicle in typical lane change situations and train SVM classifiers to predict the instant of lane change initiation with respect to the preferences of a particular driver.
We integrate this decision logic into a model predictive control (MPC) framework to create a more personalized autonomous lane change experience that satisfies safety and comfort constraints. We show the ability of the decision logic to reproduce and differentiate between two lane changing styles, and demonstrate the safety and effectiveness of the control framework through simulations.
A data-driven approach for personalized and intuitive control – Master’s Thesis (ETH)
My Master’s Thesis presents a data-driven autonomous lane change control framework that more closely models natural lane change behaviors, while still providing critical safety guarantees. The framework introduces two novel modules: a personalized lane change initiation model and an improved vehicle interaction model:
I perform experiments with a test vehicle and train classifiers to determine the instant of lane change initiation with respect to the preferences of a particular driver. This lane change initiation model allows a vehicle to autonomously change lanes in a safe but personalized fashion without the explicit initiation from the driver (e.g. activating the turn signal) or active safety features (e.g. sudden obstacle avoidance). I also train various classifiers on real lane change data to learn the most likely response of nearby vehicles to the autonomous vehicle’s lane change initiation. These two modules are integrated into a model predictive control (MPC) framework to create a data-driven, personalized autonomous lane change experience that satisfies safety and comfort constraints.
The personalized lane change initiation decision logic successfully reproduces and differentiates between the lane changing styles of different drivers. The data-driven vehicle interaction model is shown to be significantly better at predicting surrounding vehicles’ movements, and safer to use than a commonly used parametric prediction model. The safety and effectiveness of the entire integrated control framework are demonstrated in simulation. Choice of training data sets is shown to be vital for designing effective data-driven controllers, and recommendations for training data selection are provided.
Measuring Exosuit Interaction Forces – Semester Thesis (ETH)
Proper fit is of critical importance in rehabilitative robots, as poorly interfaced devices can cause blood flow occlusion and dangerous disturbance forces. Unfortunately, the consequences of an improperly attached robot often remain undetected by users who have lost sentience in those parts of their body that the device interacts most closely with. Furthermore, no reliable, purely quantitative method for classifying fit currently exists to replace the unacceptable dependence on fickle user feedback. This work proposes such a method, by using online interaction force measurements to determine whether the waist belt of the MAXX lower-limb exosuit (MAXX, Sensory Motor Systems Lab, ETH Zurich) is fitted correctly to its user.
First, the user-exosuit interaction forces created during suit actuation were measured on a testing mannequin.
Induced normal pressures were only found to occur along the hip bones, peaking at 16 kPa, and measured interaction pressures elsewhere remained under two kilopascals during various actuation simulations. A theory was developed for recognizing waist belt slip during exosuit use by observing online normal pressure readings and looking for sudden drops in pressure measurement indicating that the waist belt had slipped off the hip bone. To test this theory, a small force sensor was integrated into the waist belt directly atop the hip bone. However, further analysis showed that the normal pressures measured online during suit actuation were too small to determine suit misalignment. Instead, we noticed that in poorly fitted suits, the measured shear forces reflect the initial shifting of the suit in response to even weak actuation, and may be a viable alternative to measuring suit fit online.