Harvard, NVIDIA and Partners Release Open-Source Robot Exploration Framework RAX

In October 2026, researchers from Harvard University, NVIDIA, UC Berkeley, MIT, and Stanford University released RAX (Robot Agent eXploration), an open-source framework aimed at scaling autonomous exploration for robot agents in both real and simulated environments.

RAX uses a front-end/back-end design: the back-end handles robot drivers, simulators, and low-level controllers, while the front-end exposes only tools and skills to the agent. Given natural-language instructions, target images, and an exploration budget, the agent interacts directly with the robot, compiles successful decision sequences into standalone programs, and stores them in a persistent workspace of code, notes, and verified skills for reuse. No human teleoperation, pretrained policies, or manually designed action primitives are required, and the underlying language model parameters remain fixed.

In tests, RAX supported single-arm, dual-arm, dexterous-hand, mobile-manipulator, and humanoid platforms across multiple simulation benchmarks. As the scale of exploration increased, the agents showed continued improvement in generalization to unseen scenes; RAX outperformed the best baseline by 37.4% on average in simulation.