Stanford OnlineStanford Robotics Seminar ENGR319 | Autumn 2026 | Bridging Rigorous Methods with Machine Learning
Stanford Online
Safe autonomy framing 0:09
The talk sets up a split between classical methods and data driven methods for autonomous systems. Classical tools such as control theory, formal methods, optimization, and design give safety, stability, tractable reasoning, and proofs. Machine learning brings scale and less need for a hand built model, but its guarantees are often empirical. The aim is to bridge both so you do not have to choose between them.
Design and testing 5:00
Two paths are presented. One is to build safety and mission success by design, with hard constraints built into learning. The other is to test the system to expose when and where it fails and gather evidence for trust. The examples are framed as constrained optimal control and then as a warehouse delivery setup with aerial robots and ground carriers, no fly zones, collision avoidance, and constant operation.
Safety as value functions 17:00
Constraint value functions give a safe or unsafe test without needing you to design a new certificate from scratch. In the multi-agent case, each robot reads only local observations, yet the graph form still gives global safety. The same idea lets tiny drones carry a heavy payload while keeping distance, avoiding obstacles, and limiting wind disturbance. Trained in simulation with only a few agents, the method scaled to six, then to thousands, while keeping strong safety and task success rates.
Temporal logic decomposition 23:00
For richer missions, temporal logic describes timing, repetition, and rules such as delivery, no-fly zones, and revisits. Large language models can translate English into that logic. The logic is then broken into smaller pieces, such as reach-and-avoid or repeated tasks, until each part has a known value equation. Those atomic values are learned in parallel, then read as a graph at execution time, so the policy can satisfy the full task.
Testing as sampling 33:32
Testing is cast as a search for exogenous parameters that make cost high, by violating a safety limit or missing the goal. Traditional formal methods are robust but narrow and expensive. Blackbox learning and testing are flexible but weaker on guarantees. The goal is to find failures in complex systems without relying on a full model, while still quantifying risk and covering more than one failure mode.
Risk landscape and loop 41:32
The method turns failure discovery into sampling from a posterior distribution, using a prior belief and a likelihood built from the cost function J. MCMC can sample high-risk cases, with gradient-based or model-free versions, and has been shown to lower cost and improve robustness. An adaptive loop then tests, assesses, learns, and generates the next case, using both information gain and user utility.
Testing hidden failure modes 50:31
Large models can stand in for human reviewers and help judge whether a case is a failure worth examining. That assessment is used to train surrogate models for failure modes, then to generate test cases that cover many severities and combinations of causes. The same process can learn from real failures and invent new ones that never happened, but would break current self-driving systems. It also extends to disaster-response language models, where you compare easy and hard decisions and see how sensitive advice is to the input.
AI-generated summary. It can be wrong or incomplete - check anything that matters against the original.
