Stanford Robotics Seminar ENGR319 | Spring 2026 | Towards Trustworthy Autonomy

Reliable robotics requires a "safety-first" process that combines real-time runtime monitoring with a data-driven improvement cycle. Rather than attempting to build impossible-to-achieve perfect models, developers can create trustworthy systems by using language models to handle unforeseen scenarios and by using mathematical tools to curate training data, ensuring only high-quality information drives robot behavior.

• Semantic safety gap — traditional obstacle avoidance fails against context-based errors, such as a vehicle braking for a billboard depicting a stop sign [06:10].

• Two-stage detection — a framework uses fast embedding models for immediate anomaly detection and slower, large language models for complex decision-making during interventions [08:52].

• Hardware integration — the controller computes recovery maneuvers in parallel with reasoning tasks to ensure the robot remains dynamically stable [11:27].

• Real-world performance — quadrotors successfully identify and divert from landing zones when encountering unexpected objects, such as a keyboard or another drone [16:30].

• Black box limitations — modern end-to-end models are difficult to debug because it is unclear which training samples contribute to failure modes [18:52].

• Influence functions — a mathematical method used to trace and predict how adding or removing training data affects model success [21:26].

• Cupid algorithm — a curation tool that identifies and prunes low-quality or harmful training data, which significantly boosts success rates [25:10].

• Efficiency gains — in experimental testing, pruning two-thirds of the dataset based on influence analysis increased success rates from 40% to 90% [25:36].

• How do runtime monitors handle false positives generated by language models?

• What criteria determine if training data is flagged as low-quality during the curation process?