The research in the AMRL spans three broad themes:
Perception for long-term autonomy, including long-term vector mapping, human-in-the-loop SLAM, and robust vision-only navigation;
Failure recovery for deployed robots, including automated multi-sensor recalibration; and
Multi-agent navigation and motion planning, including high-speed multiagent navigation in adverserial domains, and time-optimal control.
Human in the Loop SLAM
Autonomous mapping of environments is challenging, especially when areas may be very large and subject to frequent minor changes. Currently, fully autonomous state-of-the-art mapping methods still struggle when faced with difficult environments and/or novice users. Using a variety of analytical and non-linear optimization tools, this project is aimed at boosting the accuracy and success rates of mapping algorithms, with minimal effort from a human.
Computationally Efficient, Safe Navigation Using Stereo Vision
This research focuses on integrated planning and perception for local obstacle avoidance using stereo RGB cameras for autonomous mobile robots. By integrating planning and perception, we expect to significantly reduce the computational requirements for safe navigation, while still remaining robust to arbitrary obstacles in the robot’s path. The test platform for this research is a Clearpath Jackal UGV.
Delta-Calibration is an automatic method for extrinsic calibration of sensors based on ego-motion. Delta-Calibration involves a closed form solution to the extrinsic calibration given necessary ego-motions performed by rigid bodies of connected sensors. An optimization based solution to Delta-Calibration can calculate even with limited axes of ego-motion, and limited information in the environment. An implementation of Delta-Calibration is available for use on our github .
Automatic State Machine Debugging
An example failure case from the Robocup Small Sized League domain. Blue lines represent the robot path, and orange lines the ball path. The left image shows a failure case in which the robot fails to transition into the kick state, while the right image shows the desired behavior.
State machines are a common tool for building controllers for various robotic tasks. The effectiveness of a state machine is often dependent on the parameters used for transition and emissions, and in many cases these parameters are a source of human introduced error. The goal of this research is to create a system which does the following:
Identifies state machines and extract their parameters from source code using static analysis.
Detects faults in these state machines via anomaly detection or domain-specific error functions.
Identifies a modification to the transition and emission parameters for the state machine that corrects the error.
Time-Optimal Control For Omnidirectional Robots
Omnidirectional robots are used in a number of domains, including service mobile robots, warehouse robots, and robot soccer. Despite their popularity, true time-optimal control of omnidirectional robots with acceleration and velocity limits remains an unsolved problem. This research aims to develop algorithms with bounded run-time to solve the time-optimal control for omnidirectional robots, while producing numerically stable solutions that may be used in iterative closed-loop control under sensing and actuation uncertainty.
Graph Planning in Dynamic Adversarial Multi-Agent Domains
One of the common approaches to motion planning in continuous space is to discretize the search space using a graph or tree and search over that reduced space. There are many approaches and decompositions that aid in constructing such graphs. In order to supplement these existing graph planning algorithms in dynamic domains with moving obstacles, we introduce strategies for supplementing probabilistic roadmaps with dynamic obstacle-dependent sub-graphs to aid in the search of high-quality paths while requiring only a coarse offline graph.
Multiple Model Learning for Ground Robots
Having an accurate motion model for a robot is equivalent to being able to accurately predict the behavior of the robot, which in turn provides the means for accurate control. Learning a motion model could be achieved through various system identification methods, which essentially train a function approximator for this purpose. However, the dynamics of a robot could not always be expressed within one single model, or in order for a single model to be complex and expressive enough, one should train it with extensive amount of training data. We approach this problem by breaking it into smaller pieces and learn multiple simple models rather than one single complex model. The robot could then switch between these models and choose the appropriate one at each time. We apply this method to the case of a ground robot and show how multiple models could be learned for different types of terrain to improve the overall motion model accuracy for the robot.