A Hybrid Meta-Learning Framework for Adaptive Safe Controller Synthesis of Dynamical(…)
summary
Proposes HyML-ASCS, a hybrid of model-based LSTM task embeddings plus gradient-based meta-learning that trains controller and barrier networks conditioned on embeddings, then runs a counterexample-guided loop with SMT-based dReal verification. Evaluated on 2D–4D benchmarks, the method cuts adaptation to a few gradient steps, typically one verify–train round, yields higher synthesis success rates and lower total computation time than prior learning baselines, and outputs formally-verified control-barrier pairs.
AI-written summary · about
chapters (12)
- 0:00Hybrid meta-learning framework intro
- 1:28Formal safety verification challenges
- 2:31Safe controller synthesis approaches
- 3:53Motivation for hybrid framework
- 4:36Model-based vs gradient-based meta-learning
- 6:02Problem formulation and barrier certificates
- 8:15Model-based meta-learner and embeddings
- 9:39Gradient-based meta-learner for adaptation
- 11:05Task adaptation and formal verification
- 13:38Experimental evaluation and benchmarks
- 14:42Conclusion and future work
- 15:44Q&A
related talks
01h 00mBetter Learning through Programming Languages
50mFlexible Neural Networks and the Frontiers of Meta-Learning
10mMIND MELD: Personalized Meta-Learning for Robot-Centric Imitation Learning
59mOn Human-robot Interactions via Multimodal Intent Modeling and Reachability-based Safety Assurance
30mAdmiral-Driven ML Framework for Marine Operations
18mActive Learning for Neurosymbolic Program Synthesis
28mGradient-Descent for Randomized Controllers under Partial Observability
48mPractical Model-based Algorithms for Reinforcement Learning and Imitation Learning, with...