Prompt Engineering for Agentic Safety Assurance in Physical AI
summary
Argues that prompts alone fail for embodied robots and proposes four pillars anchored by a system graph that encodes ODDs, mission scenarios, requirements and verification tied to ISO 26262/61508. Describes concrete tooling (Claude, Codex, draw.io, Confluence, Jira, Git), a hazard coverage skill schema that emits JSON/MD with unique IDs, and an example workflow that detects a perception model change, maps impacted scenarios, runs golden-case evals, and escalates for human sign‑off. Recommends ship-one-skill, gather traceable evidence, then scale.
AI-written summary · about
chapters (10)
- 0:00Intro and What Is Physical AI
- 0:24Why Prompt Engineering Isn't Enough
- 1:23Four Pillars and the System Graph
- 2:24Defining the ODD and Hazards
- 4:14What Belongs in a Safety System Graph
- 5:38How Agents Use the Graph
- 6:42Skill Template for Hazard Coverage
- 7:27Example Workflow: Perception Model Change
- 9:03Prompts as Context and Evaluating Skills
- 11:08Roadmap, Takeaways, and Human Accountability
