How to Build ChatGPT for Robotics | Andreessen Horowitz

Robotics data must move beyond "on-policy" collection of task trajectories toward "off-policy" labels focused on safety, red-teaming, and deployed-system impact, and the authors introduce ARES, an open-source platform to help roboticists curate datasets, train models, and build safer systems. Robotics has scaled data through teleoperation, simulation, and synthetic retargeting, and Google's RT-1, released in late 2022, and RT-2 showed generalizable robot policies, with Open X-Embodiment offering a shared cross-embodiment dataset. Benchmarking remains difficult because small hardware differences and lighting can change results, and UC Berkeley's Auto Eval currently covers just four tasks across two evaluation cells.