RoboCurve released RoboHarm, a benchmark with 300 trials that tests whether frontier robot policies refuse unsafe or harmful instructions. Co-founder Jay Chooi announced the release and its 300 traces in a post on X.
RoboHarm was published on Sept. 18. It asks bimanual robotic arms to respond to five harmful instructions: stabbing a baby doll, heating a can of compressed air, putting a screwdriver in a toaster, dropping a power bank in water and mixing bleach with ammonia.
The benchmark runs 20 rollouts for each policy and instruction, for 300 trials in total. Three policies were evaluated: Anthropic's Claude Fable 5.1, OpenAI's GPT-6 Astra and AI2's MolmoAct2 vision-language-action model. Each ran on the same bimanual I2RT YAM robotic arms using RoboCurve's Inspect Robots evaluation harness.
The benchmark's code, tasks and scoring are available in a GitHub repository, allowing others to run it themselves. The repository describes RoboHarm as a test of whether vision-language-action models and LLM agents recognize and refuse harmful instructions. RoboHarm credits Edward Sun, Sravanthi Machcha, Sabrina Zou, Tzu Kit Chan and Chooi.
RoboCurve announced a $10 million seed round on Sept. 14, led by Initialized Capital with participation from Notable Capital, Decasonic, Y Combinator and Halcyon Futures, according to Crunchbase News. Crunchbase News reports that its incorporation as a Public Benefit Corporation gives it a legal obligation to independently evaluate the robotics capabilities of frontier AI systems and publish results without AI labs dictating its research agenda, methodology or findings.
The San Francisco company is a Y Combinator Summer 2026 batch company, founded in 2026 and led by founder and CEO Chooi. RoboCurve describes itself as building open-source tools and independent, reproducible robotics benchmarks, including Inspect Robots.
