As robots are adopted and deployed in increasingly complex scenarios, simple specifications such as stability and reachability become insufficient to specify the desired behaviors of robots. Temporal logic provides a mathematical formalism for specifying complex, time-related rules. Hence, control synthesis under temporal logic specifications has received significant interest recently. This thesis focuses on a widely used logic in robotics called Signal Temporal Logic (STL), which is defined over real-valued signals. STL is equipped with both qualitative semantics, which shows whether a specification is satisfied, and quantitative semantics (also known as robustness), which measures how strongly a specification is satisfied. Taking advantage of the robustness, control synthesis from STL specifications can be formulated as an optimization problem. Traditional solutions, such as mixed integer programs and gradient-based methods, are computationally expensive (preventing real-time control), model-based (requiring the system model to be known), and centralized (for multi-agent systems).
In this thesis, we study the use of machine learning methods in STL control synthesis problems to solve the above limitations. We state our contributions in two core areas: single-agent control and multi-agent coordination.
For single-agent scenarios, our first contribution is to parameterize the control policy as a Recurrent Neural Network (RNN) so that the control depends not only on the current system state but also on the history states, which is necessary in general to satisfy STL specifications. Two training strategies for the RNN controller are proposed. The first is an imitation learning approach, where a dataset containing satisfying trajectories is generated, and then the RNN controller is trained on this dataset. The second is a Reinforcement Learning (RL) approach, where the system model is unknown and learned together with the control policy with no need for a dataset. Although these two approaches achieve very high satisfaction according to our simulations and experiments, there is no formal guarantee that the RNN controller can satisfy the specifications. Hence, we propose the third approach, where time-varying High Order Control Barrier Functions (HOCBFs) are constructed from the STL specification and integrated into the RNN controller to guarantee its correctness. Finally, in the case that the specification is not given and only a set of expert demonstrations is available, a generative adversarial imitation learning approach is proposed to simultaneously learn an STL formula describing the underlying requirements followed by the expert and an RNN control policy that satisfies these rules.
Since many real-world tasks require the collaboration of teams of robots to finish, we extend the above approaches to multi-agent coordination. We first design a novel logic called Capability Temporal Logic plus (CaTL+). CaTL+ has a two-layer STL structure designed to specify behaviors for heterogeneous teams of robots, which is more efficient and scalable than standard STL, especially when the team is large. Second, we propose a neural network framework called CatlNet to learn both the distributed control policies and communication strategies under CaTL+ specifications, showing good scalability for large robotic teams. / 2026-05-23T00:00:00Z
| Identifer | oai:union.ndltd.org:bu.edu/oai:open.bu.edu:2144/48863 |
| Date | 24 May 2024 |
| Creators | Liu, Wenliang |
| Contributors | Belta, Calin A. |
| Source Sets | Boston University |
| Language | en_US |
| Detected Language | English |
| Type | Thesis/Dissertation |
| Rights | Attribution 4.0 International, http://creativecommons.org/licenses/by/4.0/ |
Page generated in 0.0017 seconds