Verbal reinforcement learning is very effective for improving agentic performance on benchmarks, but we wanted to test how this transfers to robotics. We ran GPT 6 Astra, Gemini Argon 4, and Fable 5.1 on the LIBERO benchmark, and then iteratively developed skills which significantly improved performance. Full results linked in the comments

Frontier models like GPT-6 Astra and Gemini Argon 4 can control robots, but how much does the harness around them impact performance? We used our Nomad-Harness to test Gemini Argon on LIBERO tasks to test if text based skills could help with harder tasks that require more manipulation. We iteratively refined the skill with a critic agent reviewing failed attempts. Across 20 initial states per task, Argon’s native success rate improved: • Bowl into drawer: 0/20 → 17/20 • Mug into microwave: 0/20 → 8/20 These results show that reusable text instructions can improve robot control, and they also transfer between models, with improvements also shown for GPT-6-Astra and Fable 5.1 with the same skill.