A5073269118- Machine Learning and Algorithms
- Robot Manipulation and Learning
- Reinforcement Learning in Robotics
- Social Robot Interaction and HRI
- Evacuation and Crowd Dynamics
- Software Testing and Debugging Techniques
- Formal Methods in Verification
- Autonomous Vehicle Technology and Safety
- Advanced Software Engineering Methodologies
- Anomaly Detection Techniques and Applications
- AI-based Problem Solving and Planning
- Model Reduction and Neural Networks
- Logic, programming, and type systems
- Optical measurement and interference techniques
- Human Pose and Action Recognition
- Modeling, Simulation, and Optimization
- Advanced Vision and Imaging
- Gaussian Processes and Bayesian Inference
- Human Mobility and Location-Based Analysis
- Human-Automation Interaction and Safety
- Natural Language Processing Techniques
- Gaze Tracking and Assistive Technology
- Modular Robots and Swarm Intelligence
- Software Reliability and Analysis Research
- Robotics and Sensor-Based Localization
The University of Texas at Austin
2020-2024
Robert Bosch (United States)
2024
Robert Bosch (Netherlands)
2024
University of Massachusetts Amherst
2017-2020
Robots moving safely and in a socially compliant manner dynamic human environments is an essential benchmark for long-term robot autonomy. However, it not feasible to learn social navigation behaviors entirely the real world, as learning data-intensive, challenging make safety guarantees during training. Therefore, simulation-based benchmarks that provide abstractions are required. A framework these would need support wide variety of approaches, be extensible broad range scenarios, abstract...
We developed model-based adaptation, an approach that leverages models of software and its environment to enable automated adaptation. The goal our is build long-lasting systems can effectively adapt changes in their environment.
This work addresses the challenge of reinforcement learning with reward functions that feature highly imbalanced components in terms importance and scale. Reinforcement algorithms generally struggle to handle such effectively. Consequently, they often converge suboptimal policies favor only dominant component. For example, agents might adopt passive strategies, avoiding any action evade potentially unsafe outcomes entirely. To mitigate adverse effects functions, we introduce a curriculum...
Complex robot behaviors are often structured as state machines, where states encapsulate actions and a transition function switches between states. Since transitions depend on physical parameters, when the environment changes, roboticist has to painstakingly readjust parameters work in new environment. We present interactive SMT- based Robot Transition Repair (SRTR): instead of manually adjusting we ask identify few instances is wrong what right should be. An automated analysis 1) identifies...
We present SocialGym 2, a multi-agent navigation simulator for social robot research. Our models multiple autonomous agents, replicating real-world dynamics in complex environments, including doorways, hallways, intersections, and roundabouts. Unlike traditional simulators that concentrate on single robots with basic kinematic constraints open spaces, 2 employs reinforcement learning (MARL) to develop optimal policies diverse, dynamic environments. Built the PettingZoo MARL library Stable...
Autonomous mobile robots that use multiple depth sensors to perceive their environments, rely on extrinsic calibration combine the individual views from each sensor into a single coherent view of surroundings. Such is tedious perform manually, and requires specific scenes calibrate. Current state art automatic approaches do not consider content used for calibration, thus are robust partially informative in long-term deployments. In this paper, we present Delta-Calibration, an automated...
Accurate kinodynamic models play a crucial role in many robotics applications such as off-road navigation and high-speed driving. Many state-of-the-art approaches for learning stochastic models, however, require precise measurements of robot states labeled input/output examples, which can be hard to obtain outdoor settings due limited sensor capabilities the absence ground truth. In this work, we propose new technique neural from noisy indirect observations by performing simultaneous state...
Robot social navigation is influenced by human preferences and environment-specific scenarios such as elevators doors, thus necessitating end-user adaptability. State-of-the-art approaches to fall into two categories: model-based constraints learning-based approaches. While effective, these have fundamental limitations – require constraint parameter tuning adapt new scenarios, while reward functions, significant training data, are hard or domains with limited demonstrations.In this work, we...
Imitation Learning (IL) is a promising paradigm for teaching robots to perform novel tasks using demonstrations.Most existing approaches IL utilize neural networks (NN), however, these methods suffer from several well-known limitations: they 1) require large amounts of training data, 2) are hard interpret, and 3) refine adapt.There an emerging interest in Programmatic (PIL), which offers significant promise addressing the above limitations.In PIL, learned policy represented programming...
Action selection policies (ASPs), used to compose low-level robot skills into complex high-level tasks are commonly represented as neural networks (NNs) in the state of art. Such a paradigm, while very effective, suffers from few key problems: 1) NNs opaque user and hence not amenable verification, 2) they require significant amounts training data, 3) hard repair when domain changes. We present two insights about ASPs for robotics. First, need reason physically meaningful quantities derived...
Imitation Learning (IL) is a promising paradigm for teaching robots to perform novel tasks using demonstrations. Most existing approaches IL utilize neural networks (NN), however, these methods suffer from several well-known limitations: they 1) require large amounts of training data, 2) are hard interpret, and 3) repair adapt. There an emerging interest in programmatic imitation learning (PIL), which offers significant promise addressing the above limitations. In PIL, learned policy...
State machines are a common model for robot behaviors. Transition functions often rely on parameterized conditions to preconditions the controllers, where correct values of parameters depend factors relating environment or specific robot. In absence calibration procedures roboticist must painstakingly adjust through series trial and error experiments. this process, identifying when has taken an incorrect action, what should be done is straightforward, but finding right parameter can...
Complex robot behaviors are often structured as state machines, where states encapsulate actions and a transition function switches between states. Since transitions depend on physical parameters, when the environment changes, roboticist has to painstakingly readjust parameters work in new environment. We present interactive SMT-based Robot Transition Repair (SRTR): instead of manually adjusting we ask identify few instances is wrong what right should be. A lightweight automated analysis...
Accurate kinodynamic models play a crucial role in many robotics applications such as off-road navigation and high-speed driving. Many state-of-the-art approaches learning stochastic models, however, require precise measurements of robot states labeled input/output examples, which can be hard to obtain outdoor settings due limited sensor capabilities the absence ground truth. In this work, we propose new technique for neural from noisy indirect observations by performing simultaneous state...
Robots moving safely and in a socially compliant manner dynamic human environments is an essential benchmark for long-term robot autonomy. However, it not feasible to learn social navigation behaviors entirely the real world, as learning data-intensive, challenging make safety guarantees during training. Therefore, simulation-based benchmarks that provide abstractions are required. A framework these would need support wide variety of approaches, be extensible broad range scenarios, abstract...
Robot social navigation is influenced by human preferences and environment-specific scenarios such as elevators doors, thus necessitating end-user adaptability. State-of-the-art approaches to fall into two categories: model-based constraints learning-based approaches. While effective, these have fundamental limitations -- require constraint parameter tuning adapt new scenarios, while reward functions, significant training data, are hard or domains with limited demonstrations. In this work,...