Realset raises $10M for real-world agent data

Realset AI and Flatkey raised $10 million in Series A funding to capture real-world demonstrations and build training data and benchmarks for frontier and embodied agents, with its Household Manipulation Bench due in Q4 2026. The funding and benchmark announcement says it will test folding, loading, sorting, and wiping in real kitchens and laundry rooms, scoring each task over three trials; the archive’s agent-evaluation work shows why outside-lab success rates matter more than model demos. Builders training physical or computer-use agents should file the date: data contracts now need real-workflow coverage, not just more scraped tokens.